The Capabilities and Limitations of AI Systems at NASA

Abstract

In the past 20 years, Artificial Intelligence (AI) has made several advancements. Because of AI’s ability to process large datasets better than humans, it is thought to have a promising future in many fields. Despite their advantages, AI and specifically Machine Learning (ML) algorithms can have emergent behavior, which makes their adoption into safety-critical systems a challenge. Through an examination of the capabilities of AI systems at NASA, we see what AI is currently being used for and which algorithms are promising for future work. We also identify limitations in the potential impact of AI systems, noting that the majority of the reviewed papers focused on limitations in adopting AI systems rather than limitations in the technical abilities of AI systems. This review article provides insight into AI and ML algorithms, aviation, space and other AI-based platforms for automation.

1. Introduction

Artificial Intelligence (AI) is becoming increasingly prevalent in creating autonomous systems for space and aviation. This paper focuses on the contribution NASA has made to the capabilities and limitations of such AI systems, and it will only examine NASA-funded sources. We define AI as “a broad class of computational methods that are designed to operate with intelligence primarily by (1) learning from experience, and (2) making decisions based on learned information to achieve a goal” [1] and autonomous systems as “a system in which the computational element of the active feedback process changes in order to maintain desired performance in response to failures, threats, or a changing environment” [1]. More poetically, AI is the method by which a computer can complete tasks that were previously confined to rational minds, and autonomous systems are systems that require minimal input by humans [2,3]. We will describe capabilities as the ability of AI systems to accomplish the systems goal and limitations as the holds placed upon the systems for adoption. Understanding both the capabilities and limitations of autonomous AI systems is paramount in knowing where and how to apply these systems.

AI, sometimes called Machine Intelligence (MI), is thought to be a promising way to further multiple disciplines [4,5]—such as the need for computer vision systems that lock on to physical features while having fault detection, communication, parallel processing and extensibility [6]. ML is even proposed to be the path through which aviation safety management systems may evolve [4]. AI is better suited for certain tasks than humans, because it can process large amounts of data better than people can [4].

Unfortunately, AI and ML are confusing topics with many disagreements as to what it is, how it works and what it is capable of [7]. “The terms MI and Artificial Intelligence (AI) are often used interchangeably and describe a technology that enables a machine to (convincingly) mimic human behavior, AI being a broader definition applicable to any system, whereas MI is specific to computational mimicry. ML is a specific subset of this field characterized by algorithms that process data to build an analytical model that improves with experience (iteration), thereby ‘learning’ from datasets” [8]. It is important to note that AI only mimics human intelligence; it does not possess human intelligence. AI and ML can have emergent behavior, meaning that they are not as predictable as simple algorithms [9]. This makes their adoption in safety-critical systems a challenge. In addition, AI is computationally more expensive than simpler algorithms, creating another barrier for adoption in outer space [10].

NASA has worked on many AI systems. We have separated the AI systems broadly into the following categories: space, aviation, platforms for automation, and algorithms. Many systems will contribute to multiple categories, in which case their contributions will be listed by category. AI-based systems relevant to space travel include system management, flight path monitoring, satellite sensors, and communication.

The distinction between platforms for automation and algorithms is vague. In the context of this review, we define platforms as systems that primarily allow researchers to easily expand to future applications. Alternatively, algorithms prioritize performance over ease of use. Examples of platforms are AI and ML tools designed specifically for ease of use, such as a tool that helps improve model explainability. This is in contrast to papers that focus on algorithm creation, comparison, or improvement, which fall into the algorithm category.

For example, Rohloff 2010 [11] presents a new algorithm that used a finite-state machine to detect patterns that precede riots. In our review, we place this paper in the algorithm category because the paper introduces a new algorithm but does not make any stated effort toward ease of use. In contrast, the Spacecraft Health Inference Engine platform is purposefully designed to be easily embedded into other systems [12], resulting in it being placed into the platforms section. However, there can be overlap. For example, Dolph et al. 2018 [13] was sorted into both categories. Dolph et al. looked at the explainability of Convolutional Neural Networks using an established platform called ATTRACTOR. Because of the aim to understand the algorithm better, the study is discussed in the algorithms section. Additionally, because the paper promoted wider use through explainability, the study is also discussed in the platforms section.

This review paper will discuss each category in the following order:

• Introduction (Section 1)

• The introduction discusses the aims of this review paper, outlines the methodological approach used to collect the analyzed papers, and provides various definitions necessary to understand the following sections.

2.

Historical AI Capabilities and Limitations at NASA (1989–2005) (Section 2)

• This section outlines the growth of AI systems at NASA from 1989 to 2005 and how it was used during that time period as compared to more modern applications.

3.

Capabilities of AI and ML Systems (Section 3)

• Descriptions of AI and ML algorithms are found in this section, primarily focusing on algorithm creation, comparison, implementation and applications.

  3.1

  AI and ML Systems (Section 3.1)

  • ML algorithms can be made with a specific use case in mind, or they are tested on specific use cases. This section has subsections that cover using AI for testing on or use with specific use cases, and it also covers the creation of new algorithms and the studying of existing algorithms.

4.

Capabilites and Limiations (Section 4)

4.1

Aviation (Section 4.1)

• Aviation covers topics such as AI for navigation, control systems, simulation and various other AIs developed for the aviation industry.

4.2

Space (Section 4.2)

• The space section includes subsections on anomaly detection, schedulers, AI crew, medical AIs, AI-assisted communication, and AI for use on satellite data.

4.3

Platforms for Automation (Section 4.3)

• This section covers tools that are designed to be easily be used for different projects.

5.

Synthesis (Section 5)

• This section contains common practices and future directions identified through the literature review.

6.

Adoption Limitation—Certifications, Verification, and Validation (Section 6)

• For AI to be used in safety-critical conditions, it is imperative that they are properly tested to ensure they work as intended. This section discusses the concerns around the implementation of AI systems.

7.

Future Directions (Section 7)

8.

Conclusions (Section 8)

We do not give equal space to describe each source. This is because some of the sources are over a hundred pages of discussion, whereas others are a single-page summary, and one is even a video. In addition, some sources detail their AI systems more than others. Because of these discrepancies in length and content, we devote different amount of discussion to each source.

1.1. Aims

The aim of this review paper is to contribute an overview of AI systems at NASA between 1988 and 2024. We decided to not restrict the AI system to a singular field so that the reader might better understand the progression and future of AI systems. To this end, we have separated the sources into the historical context (Section 2), the capabilities of present AI systems (Section 3.1 and Section 4), and future directions (Section 6). While there are several reviews that discuss AI systems within particular applications [14,15,16,17], to our knowledge, a review as broad as ours (resulting from addressing an agency rather than an application) has not been published.

1.2. Paper Collection

The papers used herein all come from NASA-funded sources—primarily from the NASA Technical Reports Server. Gathering the sources from this server will ensure that the work was at least partially funded by NASA. Occasionally, the authors of the works were used to find continued work that was not published on the Technical Reports Server. However, this was only used to expand the ATTRACTOR Section 4.3.1.

The search terms that were used, and their occurrences, are shown in Figure 1. Even though autonomous systems are not the aim of this review paper, the term ‘Autonomous’ was still used as a search term because the prevalence of AI today can leave AI as an implied tool. This leads to some confusion as to whether an ‘autonomous system’ uses AI or not. In reviewing these papers, if the author states that the system is learning, then we can infer that some type of “learning agent” or AI was used. Often times, this is the only reference to AI in a paper. Thus, while the details might be elusive, AI capabilities can still be examined.

The search lead to over 10,000 papers. The abstract and conclusion of the first 50 papers for each of the five search terms were examined to see if the papers contained learning systems. About one third of these results were immediately discarded because neither the abstract nor the conclusion made a reference to a learning agent, resulting in 131 sources. The remaining were then sorted from the content of their abstracts into either history, algorithms, aviation, space, platforms for automation, or certification and validation. After this, all papers were read in their entirety, and if the autonomous system were not about learning systems, the paper was discarded. This resulted in 126 total papers. The final number of papers, with the number of sources in each section, can be seen in Figure 2.

If Figure 2 is carefully examined, it can be seen that the total number of sources across the paper comes to 150. This is larger than 126 because some sections have repeated entries from other sections. The number of sources that each section has in common with another can be found in Figure 3. The distribution of the papers across time can be seen in Figure 4. The large number of papers in 1995 can be contributed to the 1995 Goddard Conference on Space Applications of Artificial Intelligence and Emerging Information Technologies. If this conference was limited to one citation, then the $R^2$ value would be 0.500. This shows an upward trend of papers as the years progress.

1.3. Definitions

To begin talking about AI and ML algorithms, it is important to understand what these algorithms are and how they work. AI, and the phrases surrounding it, have many different definitions. These definitions can present themselves differently in different fields, times or papers. For example, the term “expert system” was popular throughout the 1990s and early 2000s to mean a system designed to perform a task that uses the knowledge of human experts. This is no longer a term commonly used to describe these systems; they are instead called AI systems or are referred to by their algorithm type. We will use the words as the authors use them within each paper. This may mean using them inconsistently or “incorrectly.” However, we believe that this represents a more realistic view of AI systems, because there are no standard specific definitions across all applications of AI systems.

Unless otherwise cited, the following definitions are directly quoted from [1]:

• Adaptive System: “The computational element of the active feedback process changes, in order to maintain desired performance, in response to failures, threats, or a changing environment.”
  Note: When intelligence is added, the system learns and adapts by adding variability. This variability is what makes it perform better and what makes it harder to certify. However, adaptive systems do not require intelligence.
• AI: “A broad class of computational methods that are designed to operate with intelligence, primarily by (1) learning from experience, and (2) making decisions based on learned information to achieve a goal.”
• Autonomy: “Systems which have a set of intelligence-based capabilities that allow it to respond to situations that were not pre-programmed or anticipated in the design (i.e., decision-based responses). Autonomous systems have a degree of self-government and self-directed behavior (with the human’s proxy for decisions)” [1,18].
• Certification: “Legal recognition by the certification authority that a product, service, organization or person complies with the requirements. Such certification comprises the activity of technically checking the product, service, organization or person and the formal recognition of compliance with the applicable requirements by issue of a certificate, license, approval or other documents as required by national laws and procedures.”
• Expert Systems: “Expert systems (ESs) attempt to capture the knowledge and experience of human experts and then use this information to form rules that suggest the best response for a given state of the world. Historically, expert systems were one of the first successful methods of applied artificial intelligence.”
• Fuzzy Logic: “In classical set theory, the membership status of an element is binary—it is either a member of the set, or it is not. With a fuzzy set, however, elements can have a partial ‘degree of membership’ within the set. The extent to which an element is in the set is defined by a membership function, which is valued in the interval [0,1].”
• Machine Learning (ML): “ML, a subdiscipline of AI, involves the development of prediction or decision algorithms not explicitly programmed to predict or decide but rather that learn from data representing past predictions or decisions” [19].
• Nondeterminism: “In general, a nondeterministic algorithm is one in which the output cannot be repeated with certainty, given the same input. …[There are] four types of nondeterminism: Environmental nondeterminism, probabilistic algorithms, uncertain existence of solutions, and concurrency.”
  Note: While determinism is not a requirement of NASA software, it is highly encouraged.
• Probabilistic algorithms: “This includes algorithms that are based on sampling a random process or probability distribution. Mathematical techniques to bound the behavior of these algorithms and prove their convergence would be necessary if they were to be used in a certified system”
• Qualitative Physics: “Qualitative physics is a form of knowledge representation designed to reason about physical systems using only qualitative abstractions. Essentially, the model of the world is expressed at an abstract level using qualitative attributes of behavior rather than numerical simulations.”

2. Historical AI Capabilities and Limitations at NASA (1989–2005)

To best understand where the future of the capabilities and limitations of AI systems leads, we must examine the change over time to gauge the rate of growth. When AI was beginning, it required an expert to be adequately used; so in order to apply ML, you would have to be an expert in the field that the data came from, data processing, and AI [20]. This is too much to expect from one person, and so NASA made many AI systems that could circumvent the need to be an expert in data processing and ML. Despite the challenges associated with making MI systems, NASA had success at developing several of these systems in the late 1980s and early 1990s. There is some difficulty in comparing how effective these early systems are in contrast to modern systems, because they often do not use the same evaluation metrics that are now commonplace. Also, authors sometimes only simply state if a system worked sufficiently well to meet their goals. However, the scope of the projects can often be used to gauge progression.

Early systems still hold many accomplishments, including the quick and automatic detection of peculiarities in engine behavior with an AI and neural net architecture [21], an AI-enhanced service manual for a hazardous gas detection system [22], and a real-time neural network to diagnose jet and rocket engines [23] as early as 1989. At the Automation and Robotics for Space-Based Systems–1991 conference, many authors presented intelligent systems, including robotic arm positioning aided by AI, neural networks for modeling the impulse response of systems, an adaptive neural network for autonomous robot control, and a preliminary look at a machine vision guidance system for the automated construction of space structures [24].

By the time of the 1995 Goddard Conference on Space Applications of AI and the Emerging Information Technologies, AI was much more capable than in the early 1990s. AI was able to accomplish increasingly complex tasks, including the following: make basic predictions of the space environment with better visualization [25], create small databases from regular language [26], update and provide rules for a database for non-experts via supervised training with expert humans [27], quickly allocate resources using a genetic algorithm [28], control a robot and spacecraft using a neural net [29], detect star clumps with a generic algorithm [30], predict the outcomes of high-resolution radiometers using linear and non-linear prediction [31], and begin monitoring spacecraft health with fuzzy logic [32]. Outside of this conference, AI was used to point spacecraft at the desired target so that better images could be formed; this same algorithm was also able to perform mosaics [33]. In 1996, NASA began work on combining fuzzy logic with neural networks [34]. At the time, the neural networks had about 50 total nodes split across five layers [34]. The problems solved by AI in 1995–1996 are more complex than the problems shown in the reports from 1989 to 1991. AI in 1995–1996 was capable of interpreting human language, performing basic computer vision, and supervising database management. Genetic algorithms and fuzzy logic saw more utilization and better performance in the mid 1990s than the early 1990s. This increase in output came from a confluence of multiple reasons, including increased compute abilities, more efficient algorithms and the use of different algorithms.

In 1997, the first spacecraft to feature an onboard AI-enhanced planner was being built [35]. This planner reduced cost because it could adapt to unexpected fortunate events [35]. In 1998, NASA began working on making a neural network where each neuron is a reinforcement model in order to route internet traffic [36]. In 1999, algorithms were developed that greatly sped up the image processing of synthetic aperture radar [37], and that enhanced the deep space tracking of spacecraft [38]. In 2000, an AI planning and scheduling technology was developed to configure and control interferometers in order to speed up usage so that more scientists might use the interferometers [39]. In addition to this, an algorithm was developed to control spacecraft with natural commands such as ‘during the next week take pictures of the following asteroids and thrust 90% of the time’ [40]. In 2002, an algorithm was developed to reduce the landing space for a Mars lander by a factor of 10 and had an accuracy of 96% [41]. These systems are more capable than their earlier counterparts. In the late 1990s and early 2000s, there was enough trust and confidence in AI systems that it was utilized on board a spacecraft that launched in 1998. This shows an improvement in the capabilities of AI because of the rigorous testing and validation that NASA requires [42]. In addition to this, more work was being undertaken to improve neural networks, and great speed improvements were made in the early 2000s.

From the early 1990s to the mid 2000s, a great amount of improvement can be seen both in the capability of AI systems and in the complexity of problems that can be solved by these intelligent systems. For example, in the span between 1989 and 2005, there were many improvements to ML and processing power. In 1996, a neural network was trained that had 2, 14, 49, 7 and 1 neuron in its five layers, and six years later, a neural network was trained with 20 layers with 10 neurons per layer [34,41]. In the early stages of AI, NASA was focused both on making AI accessible to researchers who did not have an expertise in ML and on applying the algorithms to space-centered problems. As ML became more commonplace, NASA took careful consideration over how to validate such systems [20,42].

3. Capabilities of AI and ML Systems

3.1. AI and ML Systems

To use AI effectively requires skill [5]. AI and ML ought to be made with knowledge in the application domain, and in some cases, their outputs must be interpretable by humans [4]. In order to make more effective use of ML/AI resources, interdisciplinary fields need to have ontological standardization (standard terms and descriptions of experiments, data, and findings) [8]. Building AI is also difficult because the user feels performance in terms of response rate and perception of correctness [43]. For example, if a developer wants to make an AI that will be adopted for use, they are required to know what makes an AI “feel correct.” This is a soft requirement that is not necessary for the function of the algorithm, but it has the possibility to determine if the end user is comfortable using the AI.

This section covers various applications of known AI systems and NASA’s contributions to algorithm development and understanding. Some papers do not discuss the models in depth and only make vague allusions to learning algorithms or “well-known AI principles” [44]. There are many reasons to withhold what particular AI models are being used. The models could be proprietary or a re-use of a previously documented model. These papers are good for obtaining a notion of what AI is being used for; however, they do not allow us to track specific improvements or changes in models. This section aims at covering a wide variety of applications from 2008 to 2024, focusing on tracking AI improvements, changes, and performance. Thus, we will only include sources that document what type of AI is used in this subsection. Even among the documented AIs, many papers leave out specific performance metrics and only declare if it worked for their specific use case. This makes a direct comparison difficult. However, these papers are still included, because they show trends of how AI is used throughout time.

3.1.1. Applications

This subsection covers a wide variety of applications from 2010 to 2024, focusing mainly on the performance of algorithms in an application.

In 2010, NASA developed a new sequential pattern methodology to identify patterns that precede riots, rebellions, and coups [11]. This model was created using a finite-state machine, and it was shown to have a low false-alarm rate. Exact metrics were not discussed; however, it performed well in the test case.

By 2014, NASA was using a Bayes network for sense-and-avoid conflict resolution [45]. This network can determine which direction to turn the UAS to avoid threats and collisions. The work was preliminary and did not yet have results; however, it is a step in the march toward autonomous aircraft. That same year, NASA created a neural network with heavily weighted safety concerns to ensure that safety violations do not occur while in flight [46]. The algorithm was safe and reliable while remaining comparatively quick and achieved 92% of the theoretically optimum performance. NASA also created a Multi-Layer Perceptron (MLP) for the simulation performance classification of a physics-based aerospace vehicle simulation given a series of inputs [47]. The output layer of the MLP was reduced with a Support Vector Machine (SVM) to give the simulation a pass–fail prediction. This algorithm can be used to speed up simulations because it can suggest to omit a simulation that is likely to fail. The MLP successfully predicted if a simulation will result in a pass an average of 80% of the time with ten training instances.

In 2015, NASA released a paper going over friction–stir–welding properties [48]. The features are spindle speed, plunge force, welding speed, and an empirical force index, and the predicted outcome is the ultimate tensile strength of aluminum alloy joints. They used an adaptive neuro-fuzzy inference system (ANFIS) and an ANN. The ANN used is an MLP trained with back propagation. The evaluation metrics used are Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). The ANFIS architecture is based on human knowledge and uses a neural network as its backend, where human knowledge informs the weights attributed to each node. Both models were trained using leave-one-out cross-validation. In total, 73 unique weld types were made by a machine and were then fed into the two different models. The ANN (w/o spindle speed) produced an of RMSE 36.7 MPa and an MAPE of 10.09%, while ANFIS (w/o spindle speed) produced an RMSE of 29.7 MPA and an MAPE of 7.7%. Both are good results with ANFIS outperforming the ANN. Also in 2015, a supervised-autonomy robotics platform was developed [49]. This platform uses a camera, laser scanner and inertial measurement unit to create a 3D map of the environment. They used this with a unique ML algorithm to detect obstacles. These data were used with a nearest-neighbor algorithm to detect its own poses. The platform also had enhanced planning and was able to perform complex movements from simple commands from the human operator (including movement and pathing).

One problem addressed in 2016 was reducing the simulation data required for knowledge bots that determine if a simulation will result in a favorable outcome [50]. The authors of [50] attempted to find the decision manifold to reduce the number of simulations required. They used Reinforcement Learning for feature extraction and an SVM for the separation of features, minimizing the number of simulations that are required to train the DNN from 10,000 simulations to 25.

In 2017, researchers created an Autonomy Operating System (AOS) for UAVs featuring diagnostic reasoning, planning and natural language processing [51]. This was completed using several ML methods including Bayesian reasoning and a non-descript “Machine Learning” tool.

In 2018, the Pilot Engaged Expert Response (PEER) system used a re-trained Watson Discovery Advisor (WDA) to give pilots and crew relevant information to assist in their duties while in the air [52]. WDA is a complex model featuring natural language processing and ML to find information in unstructured text with simple prompts by the user. WDA uses five steps to give the answers: (1) natural language processing of the question, (2) text retrieval queries that find sections of the text that may respond to the question, (3) natural language processing to determine which queries are most relevant to the question, (4) ranking the responses, and (5) displaying the top result. This system was shown to correctly answer questions with a notable example correctly identifying why wind speed was being displayed erratically. Another effort in 2018 involved automating the process of generating air traffic scenarios [53]. The automatic scenario creation speeds up data collection for automated systems, and it has validation techniques to ensure that the quality of generated scenarios matches the quality of hand-designed scenarios. A final example from 2018 shows how a convolutional neural network was developed to detect optical fiber placement defects on a per-pixel basis [54]. A Resnet architecture was used, wherein the hyperparameters were tuned using a genetic algorithm. The authors noted problems obtaining a large enough balanced dataset to achieve great results. Their results were promising; most faults were identified as a fault, but the specific fault type was often mislabeled.

In 2019, Minnis et al. used an ANN to detect multi-layer ice-over-water clouds in satellite data [55]. Each hidden layer of the ANN had 50–70 neurons. They achieved a combined multi-layer and single-layer cloud hit rate of 80% in the day and 77% at night. Minnis et al. reported a cloud layer height that was accurate to ±1.6 km and an optical depth accurate to ±0.556. In another application, NASA introduced a reinforcement algorithm for adaptive stress testing that finds the most likely path for a failure [56]. The problem is set up as a Markov decision process, and Reinforcement Learning is used for optimization. In large systems, an exhaustive search for potential failures becomes difficult and “generally cannot prove the absence of failures” [56]. This system examined multiple simulations where the learning algorithm had (1) full access, (2) partial access, and (3) no access to the simulation state. They used a modified Monte Carlo algorithm to generate simulation states and examined how well their system worked in an airborne collision system. The algorithm performed as expected and helped to find faults.

Tong (2020) showed a supervised ML to help in the design of turbofan engines by predicting their performance specifically in terms of cruise thrust-specific fuel consumption (TSFC) and core sizes [57]. This was accomplished using DNNs for cruise TSFC and SVMs for core size. They did an 80–20 train–test split and implemented 6-fold cross-validation. Both models achieved an accuracy of ∼98%.

In 2021, NASA worked on Gaussian Naive Bayes, SVM and ANN to identify rats that are more likely to be impaired after being exposed to galactic cosmic radiation [58]. To categorize the rats as impaired or not, statistical methods were applied to a control group of rats to determine a threshold for performance. An attentional set-shifting test was then performed on the rats before the radiation exposure. The results from this test and the dosage of radiation are used as inputs to the ML algorithms. A variety of ML algorithms were chosen because they employ different mathematical concepts as part of training and inference. Overall, 95% of the rats were unimpaired, resulting in an imbalanced dataset; class weighting was used to account for this. The result was mixed but showed more prediction power than random chance: between 50% and 86%, depending upon the model and radiation dosage.

Recognizing that ML is difficult to implement correctly, this team released a tool to help others utilize ML appropriately in 2022 [59]. This tool created a pipeline—from pre-processing to model deployment—to help those who are less familiar with AI be able to implement AI systems. Also in 2022, NASA used satellite data from China to fill in the surface ${\mathrm{NO}}_2$ concentration [60]. They integrated weighted space and time information with an extra-trees and deep forest model to fill in the missing data. The model achieved an R-squared value of 0.89–0.96, showing that it has good performance on the test set. NASA also used satellite data from low earth orbit and geostationary earth orbit satellites along with a CNN to generate high spatiotemporal resolution temperature data [61], reporting an MAE of 1.73 K. The CNN was called NASA Earth eXchange Artificial Intelligence Land Surface Temperature (NEXAI-LST). The combined data performed better than just GEO or LEO. In 2022, NASA also worked on precipitation type (convective, stratiform, mixture, no precipitation, and other precipitation) classification using satellite data [62]. A dual-frequency precipitation radar was used as ground truth for training the ML models. Five out of the six models they tried were able to achieve 90% or greater accuracy (SVM, Neural Network (NN), Random Forest (RF), Gradient Boosting, and 1D CNN). All variables were normalized and unitless before training. To account for class imbalance in the data, the authors used (1) a random majority undersampling method to randomly removes examples from the larger classes and (2) a smoothed bootstrap resampling technique to generate synthetic examples of low-count classes.

The year 2023 saw significant increases in ML applications. In [63], NASA produced an ML model for predicting a turbofan’s TSFC, engine weight, core size, and turbomachinery stage counts using inputs for a turbofan design. The ML algorithm was implemented using supervised deep learning and K-nearest neighbors regression. This algorithm achieved 90–98% accuracy depending on which parameter is being predicted. In [64], the authors used AI to monitor major Brazilian aquifers with Gravity Recovery and Climate Experiment (GRACE) data (with wavelet and seasonal decompositions) and ground-based hydrological measurements. The AI approaches examined were Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Model and CatBoost, which were followed by a linear regression step. Satellite data from GRACE were used as input for the ML models. Linear regression worked the best, outperforming Global Land Data Assimilation System (GLDAS) estimations. In [8], multiple ML algorithms (SVM, RF, XGBoost and logistic regression) were used to detect small-cell lung cancer in CT scans. This study also looked into using a CNN for cancer risk assessment to see how aggressive the treatment plan should be. Performance showed good results and the algorithms demonstrated potential for both detection and risk assessment. In [65], the authors used a deep Recurrent Neural Network (RNN) to determine the feedback sensitivities between permafrost degradation and release of soil carbon. In total, 13.1 million unique field measurements were used to train the model. These data also had high dimensionality, and so they used various dimensionality reduction techniques to reduce it to only 56 features. The RNN had 2.5 million parameters and was called GeoCryoAI. In [66], the authors describe the impact AI has on protein modeling. They found that AI can help to rapidly evaluate protein engineering tests. Sequence-based feature embedding (derived from NLP) seems to be the most popular way of embedding protein features. NLP does so well because proteins can be represented as a string of amino acids, and NLP works well with strings of data. Small labeled datasets are common in the field of protein modeling, and so simple models such as SVMs are used instead of larger NNs that require more data. In [67], NASA made an MLP and RNN to predict the stiffness and fatigue life of an eight-ply symmetric polymer metric composite laminate. Both models performed well, though RNNs performed better at predicting the fatigue.

In 2024, NASA released a video showing a chat bot powered by ChatGPT-4o that interfaces between a graph database and an end user for air traffic management [68]. System engineers report completing their tasks up to seven times faster using the chat bot. The chat bot even suggested some requirements that the engineers did not initially think of.

3.1.2. Summary and Synthesis of Applications

A total of 26 papers from 2010 to 2024 were described in this applications section; their distribution timeline can be seen in

Table 1. This table shows the distribution of the years of the papers used in Section 3.1.1 of Section 3.1.

Year	Number of Papers
2010	1
2014	3
2015	2
2016	1
2017	1
2018	3
2019	1
2020	2
2021	1
2022	4
2023	6
2024	1

. In 2010–2015, AI was designed to detect riots, resolve mid-flight conflicts while maintaining safety, and predict the properties of friction stir welding.

In 2016–2020, NASA used AI to reduce the simulation data required to train an AI algorithm, create an AI-enhanced UAV operating system, perceive and interact with the world in a robotics context, provide relevant information to the crew of an aircraft, create air traffic scenarios, detect optical fiber defects, detect multi-layer ice-over-water clouds, find likely routes toward failure in stress testing, and design turbofans.

The years 2021–2024 showed additional progress, as NASA used AI to identify rats that are likely to be impaired from exposure to cosmic radiation, create a platform which allows non-experts to easily use AI, detect surface ${\mathrm{NO}}_2$ from satellite data, generate surface temperature from satellite data, classify precipitation types, improve turbofan designs, monitor aquifers, detect lung cancer, find sensitivities between permafrost degradation and the release of soil carbon, model proteins, predict the stiffness and fatigue life of a polymer composite laminate, and create requirements for air traffic management.

NASA has tackled more complex problems with AI as the years passed. The increase in complexity and dataset size (such as those produced by satellites) shows how AI improved over the past decade. That being said, recent improvements are not at the rate as was seen from the 1990s to the mid-2000s.

Over the past couple of decades, the use of neural networks has increased. This could, in part, be because of an increase in computational power over the past decade or the good performance of neural networks across a wide variety of use cases. Neural networks, and other AI algorithms, need good data to perform well. Well-balanced large datasets were identified as a necessary step forward for the improvement of AI techniques.

The open research for this subsection is to find new application spaces or researchers applying different algorithms to existing application spaces. We recognize this is a broad recommendation. However, due to the broad nature of applications, it is difficult to provide specific recommendations.

3.1.3. New Algorithms and Insights

NASA has worked to create new algorithms, improve existing ones and gain insight into ones that already exist. The improvements can come in many forms such as combining multiple algorithms or creating a new objective function. This section provides an overview of NASA’s progress in this area from 2008 to 2022.

In 2008, NASA produced an ANN for autonomous real-time moving object detection with changing illumination and perspective [69]. The new improvement they made was combining two different object-recognition methods: one based on shape and one based on color.

In 2009, NASA created an evolvable synthetic neural system that included at least one neural basis function, where each of these neural basis functions included an “evolvable neural interface operably coupled to a heuristic neural system to perform high-level functions and an autonomic neural system to perform low-level functions” [70]. These evolvable systems can be combined to create an evolvable system hierarchy. That same year, NASA released A Complexity Metric for Automated Separation [71], wherein they define a Maneuver Option (MO) that is a metric for determining the directions an aircraft can move that do not cause conflict within five minutes. The directions considered are 15° and 30° left and right, straight, and five different altitudes at 1000 feet increments. The newly developed metric was able to quantify the complexity of an airspace at a given moment in time and then identify aircraft that were particularly at risk. While not directly an AI algorithm, the MO metric development did introduce a level of automation in air traffic monitoring.

In 2010, NASA introduced a sequential pattern methodology used to identify patterns that precede riots, rebellions and coups [11]. The sequential pattern methodology “generates easily interpretable patterns based on direct observations of sampled factor data for a deeper understanding of societal behaviors that is tolerant of observation noise and missing data” [11]. This new algorithm is designed to be best at forecasting future events, and it showed promise in predicting riots and rebellions.

NASA patented a new ML algorithm, modeled after the immune system, for aircraft defensive maneuvers in 2014 [72]. The same year, NASA created a “memetric learning” ML model [73]. In memetric learning, memes are shared across multiple agents. The agents use the memes as a set of instructions to complete their tasks. The best memes are then kept and new memes are made from these memes. This idea is similar to RL and GA, and it is intended to be less brittle to new environments than typical AI.

By 2016, ML was shown to be a feasible and effective way of quickly modeling simulations [50]. NASA also created a new mathematical representation of the dispatchability property of a simple temporal network [74].

In 2018, NASA tackled the problem of human speech variability and background noise making speech recognition difficult. They determined that the difficulty can partially be mitigated by choosing an appropriate vocabulary consisting of sounds that were “different enough” for a machine to effectively separate the selection [75]. This paper also presents various other errors that prevent making a good speech recognition system, but it did not address them. Also in 2018, Dolph et al. looked into what features a CNN locks onto during object classification by looking at the intermittent layers [13]. Early layers are image primitives similar to edge detectors, while the higher layers are more abstract. At those higher layers, deconvolution can be used to gain some visual idea of what is being learned by the CNN. In another work that year, Zhang et al. introduced a new algorithm for a self-organizing map (SOM) [76]. SOMs are ANNs used to “map” a high-dimensional input to a low-dimensional space while preserving the topology relations [76]. Self-organizing maps are typically random; however, this paper proposed a way to make a deterministic model. This process can also be applied to other AI algorithms. To ensure their process worked, they tested it on cloud classification. Their SOM algorithm outperformed a k-means algorithm and produced quality cloud regimes.

By 2020, NASA had completed a formal approach to generating runtime monitors for UASs in a structured natural language [77]. FRET (Formal Requirements Elicitation Tool) allows the user to input structured natural language to output requirements for the UAS that can then be set as hard requirements in a runtime verification software Copilot (v4.2). Copilot monitors the system in real time to ensure that the requirements are not broken. This entire system is called the Independent Configurable Architecture for the Reliable Operations of Unmanned Systems (ICAROUS).

In 2021, Cooper et al. used an external perception system as a way to detect a clutter field during navigation [78]. Rather than detect individual items to move around, a clutter field allows an algorithm to focus on the macro level, creating zones with different clutter levels so that multiple autonomous agents can stay within their given maximum clutter level.

In 2022, Razavi et al. presented an argument and method for merging ML algorithms with the historic standard of Process-Based Modeling (PBM) [79]. The current framework for the hybridization of ML and PBM has three parts:

• Sequential framework—ML receives the output of a PBM and estimates error or model uncertainty.
• Modular framework—ML is a sub-model for PBM that can serve as an analog for processes that are not yet well understood, or PBM can serve as an input to ML when the process is well understood.
• Surrogate modeling—ML is used to make an emulator of some PBM with a better computational efficiency.

Razavi et al. argues this is not enough, and that co-evolution is necessary to receive the advantages of both. Examples include designing a training function that enforces physical relationships with known differential equations or using ML to derive equations.

3.1.4. Summary and Synthesis of New Algorithms and Insights

A total of 14 papers from 2008 to 2022 were used; their year distribution can be seen in

Table 2. This table shows the distribution of the years of the papers used in Section 3.1.3 of Section 3.1.

Year	Number of Papers
2008	1
2009	2
2010	1
2014	2
2016	2
2018	3
2020	1
2021	1
2022	1

.

In 2008–2015, NASA created a real-time moving object detector that locked onto both shape and color, an evolvable synthetic neural system, a complexity metric for automated separation, a sequential pattern methodology to forecast future events, an ML model designed after the immune system, and an ML model designed after memes.

In 2016–2020, NASA created a new mathematical representation of the dispatchability property of a simple temporal network, a deterministic algorithm for a self-organizing map, and a formal approach to generating runtime monitors for UASs in a structured natural language. In this same time period, NASA was looking into how the algorithms work and discovered neural networks were starting to outperform other algorithms in some applications, vocabulary selection can make speech recognition easier or harder, and what features a CNN prefers at what layer of convolution.

In 2021–2022, NASA created an algorithm for clutter field detection to aid in robotic navigation, and they considered how to best use AI in tandem with more traditional PBM methods.

As time progressed, NASA seemed to shift focus from creating new AI algorithms to finding new insights about existing algorithms. Understanding how the algorithms work on a deeper level can help reduce the black box nature of the more effective neural networks and allow people to more fully understand their limitations. In addition, NASA is working on making AI tools that can work in tandem with physics-based models to create AI that can behave more in line with the real world. These adjustments will make AI more reliable and trustworthy.

Despite these advances, there are still open research challenges, including the following:

• Understanding how “black box” models work.
• Creating explanations for how AI algorithms work.
• Creating AI models that use domain expertise, such as PBM, to make predictions.

Most of the future directions have a common theme of “understanding AI systems.” This is because having a better understanding of AI systems can lead to better performing algorithms and a wider adoption of AI systems. In order for the algorithms to be more widely adopted, the system must be more than just easier to understand; it must also be explained to the end user.

As can be seen in

Table 3. This table shows the capabilities and limitations of Section 3.1.

Capabilities	Limitations
AI performs well in a wide variety of application space.	Not all AI has been tested for all application space.
We are beginning to understand “black box” models.	“Black box” models are still not well understood.
Domain expertise is being blended with AI systems to provide a more understandable output.

, AI has many capabilities. Because of its adaptability, it performs well on a multitude of tasks. The main challenge addressed by the reviewed papers is in adoption. If the systems do not meet performance expectations, satisfy certification requirements, or produce explainable outputs, then they will not be used.

4. Capabilities and Limitations

This section will outline the capabilities and limitations of AI algorithms, which are broadly divided into three main categories: aviation, space, and platforms for automation. The state of the literature for each of these categories will be discussed, and open challenges will be identified. It is important to note that by capabilities, we mean the ability of the system to perform its given task, and we focus on limitations that prevent adoption. While there are other types of limitations (for example, technical limitations, operational limitations, and regulatory challenges), the main hindrances identified by our review are those to adoption.

4.1. Aviation

A large portion of the literature within the aviation category is related to AI-enhanced Air Traffic Control (ATC). Airspace around airports is complex, and ATC systems must consider the many different types of aircraft present as well as the large volume of traffic [80]. The complexity was expected to surpass human limitations by 2024 [81]. Because of this increase in complexity, there is a need for AI-enhanced ATC [81,82,83] that matches the safety of current ATC systems [82]. There are several ways of mitigating safety risks; one common method is to have AI handle common, day-to-day scenarios and notify a human operator of situations that are unusual or complicated [81,82].

In addition to managing piloted air traffic, ATC systems must now account for unmanned aerial systems. Such unmanned systems are utilized in many applications, including military, search and rescue, traffic monitoring and reporting, wildlife monitoring and surveys, fire and flood monitoring, pipeline and transmission line inspection, aerial photography, crop dusting, and package delivery [45]. On top of this, the complicated nature of UAVs and ATC makes users reticent to adopt AI technology [7,81]. This section (Aviation) covers the breadth of AI used at NASA for aviation, and it includes subsections of navigation, system control, simulation, and others, ending with an overview of open challenges in the field.

4.1.1. Navigation

In 2009, NASA released A Complexity Metric for Automated Separation that discussed metrics to determine airspace complexity [71]. The metrics are functions of the ability of aircraft in a region to safely make direction changes. In this work, Aweiss proposed two such metrics that could be calculated for a region of airspace: one is the number of aircraft that cannot change trajectory, and the other is the average number of possible trajectory changes available for each aircraft. Both metrics could be used as an objective function in ML, allowing them to aid in the creation of an automated ATC system. For example, an ML algorithm could try to minimize the number of aircraft that cannot change trajectory or try to maximize the average number of possible trajectory changes for each aircraft. In the same year, NASA also released a tech brief that presented the development of an algorithm for the flight path planning of autonomous robotic blimps [84]. These algorithms were developed within a general framework, allowing them to be applied to submarines and other surveying robots in addition to their original intent of flight path monitoring. The algorithms were also wrapped in a user-friendly interface, allowing them to optimize various path-related variables, such as distance or time.

By the mid 2010s, NASA had designed a flight-planning algorithm to minimize the noise pollution of rotocrafts (small drones with one or more overhead propellers) affecting nearby communities [85]. Rotocrafts are of interest to NASA because they generally operate independently and do not interfere with fixed-winged airports. The local search algorithm they employed was a simplistic model, but it was able to find quieter flight paths than the standard way [85]. NASA also focused more directly on ATC efforts in this decade [81]. While AI-enhanced ATC helps with day-to-day operations, the users of the tools had difficulty trusting the AI. They often inhibited the AI from working—especially in time-sensitive situations [81]. In addition to human-based ATC, NASA investigated the use of AI tools to enable sense-and-avoid systems in unmanned aircraft [45]. Barreiro et al. presented an intelligent sense-and-avoid algorithm that automatically uses aircraft type, transponder code, communications, local routes, airports, airspace terrain, and weather to predict the flight path of the surrounding aircraft [45]. This software is capable of detecting conflicts and resolving them with probabilistic path planning. This in turn allows for better ground control because of the more accurate flight path information.

In 2018, NASA released a paper that looks at the air traffic around the Dallas–Fort Worth area. Bosson et al. built upon previous work, this time including wind concerns [80]. The wind can cause superfluous alerts and a delay in conflict resolution. The algorithms—Advanced Airspace Concept and the Terminal Advanced Airspace Concept—aim to ensure the separation of aircraft and efficient departures and arrivals. That same year, NASA also released Automated Scenario Generation for Human-in-the-Loop Simulations [53], which aims at automating the previously manual process of creating the initial human-in-the-loop scenario for AI-enhanced ATC. The humans tended to keep many of the characteristics created by the AI. The impacts of this simulation will be listed in Section 4.1.3.

NASA also made an AI to give information to an aircraft’s onboard crew, enabling them to fly more safely [86]. The Autonomous Operating System (AOS) they developed does not operate completely independently from ATC, but it does have the ability to communicate with them via natural language. This platform is meant to be an operating system that is modular, being similar to an app on a phone. AOS can perform basic pilot procedures and has automated reasoning capabilities for procedures that require more thought. AOS achieved variable autonomy because it can be flown with direction from ATC. In 2020, Fernandes et al. discussed a program that alters the flight plan both before take off and during flight to reduce cost and respond to the uncertainties found in airspace [87]. At the time of the paper’s release, mid-air flight path changes were made verbally. The verbal process is manual and slow, and it might not be most optimal for the entirety of airspace. The computer program suggested in [87] would be necessary for the timely FAA transition to trajectory-based operations.

4.1.2. System Control

In 2014, NASA filed a patent that showed a method for controlling one aircraft relative to another [72]. The parent aircraft can either move constantly or take evasive actions, and the child will still move to the desired location and orientation relative to the parent [72]. This has combat applications and also applications to congested airspace. In 2017, NASA released Autonomy Operating System for UAVs: Pilot-in-a-Box that first introduced the AOS mentioned in the previous section [51]. AOS draws inspiration from the long use of iOS and Android to create an environment where creating apps is easier for the developers. It comes with a built-in “pilot in a box” which allows for speech communication with ATC, and some general purpose AI for diagnostic reasoning, plan and procedure execution, and natural language processing.

A 2018 contractor report provides an overview of the PEER system [52]. The aim of the PEER system is to act as an AI-enhanced advisor capable of providing useful information to the flight crew in real time. At the time of publication, PEER was still in development. The researchers made several advancements, including making a question-answer training set pipeline which increased PEER’s ability to answer questions. The goal was to place PEER onto an aircraft to provide relevant information to the pilot. The PEER system takes into account the fact that automation can lead to people becoming lax, and its aim is to enable pilots to learn and improve their skills, being a peer that aids instead of a co-pilot that takes over.

Dutle et al. developed an approach to creating monitors for AI systems from natural language [77]. The approach is built with the intention of enabling autonomous UASs in airspace. The algorithm accomplishes this by doing runtime verification that detects when rules provided by natural language have been violated. An example of a natural language command is: “While flying, remain separated from an intruder aircraft by at least 250 ft horizontally or 50 ft vertically.” In 2023, NASA released another paper that implemented a natural language processor, this time focused on written documents [88]. The algorithm was capable of extracting flight operation procedures and constraints from air traffic management documents with a 96.6% accuracy. The software converted the natural language to constraints such as altitude, speed, position, trajectory heading, and arrival time. Language-to-constraint translation works well because of the precise and consistent language used in aviation.

4.1.3. Simulation

In 2013, aviation systems were beginning to be autonomous in the air, but not from the ground. Rungta et al. looked at a simulation using both AI and humans to create validation techniques for the safety of AI-human interaction [82]. Air traffic, and other dynamic traffic environments, can have changing objectives. This can make path planning a computationally intensive endeavor. Colby et al. discussed a cooperative co-evolutionary algorithm to create a multi-agent controller while reducing computational load [46]. As UASs become more prevalent, it is important to allow these systems to dynamically obtain information and change their flight paths. The work in [46] also shows that a cooperative co-evolutionary algorithm can create a multi-agent dynamic flight path while ensuring no safety rules are broken. The algorithm outperforms deterministic search algorithms in terms of computational speed while achieving 92% of the optimal solution.

In 2020, NASA released multiple papers regarding simulations of airspace scenarios. Oza et al. looked into detecting safety-relevant anomalies with anomaly detection, active learning, and precursor identification to try to detect errors even before they happen [89]. They showed the teamwork between human experts and AI to detect and identify the precursors. The main problem was preparing the dataset to be properly balanced, because anomalies are, by definition, rare. If they left a realistic balance, then the ML algorithm would not lock onto the anomaly and would instead say everything is normal. However, their simulations still showed promising results for anomaly detection of UAV swarms.

The next two papers came out of the Baculi group and discuss AI-enhanced fire detection [90,91]. AI-enhanced UAVs can detect fire fronts, freeing people from manually checking the fires and making the process safer. UAVs can be quicker, safer, and more cost effective than manually flown UAVs. Fire Front Detection and Tracking for Autonomous sUAS in STEReO goes over a proof-of-concept simulation for UAV-based fire detection [90]. This simulation uses different fire scenarios combined with different algorithms and enables the detection of a dynamic fire. This simulation enabled the physical creation of the autonomous sUAV for fire detection shown in Towards An Autonomous sUAS Operating in UTM TCL4+ and STEReO Fire Scenario [91]. Gen 2.1 of the physical model is equipped with lidar, IMU and computer for data processing, Gen 3 adds a range finder altimeter, downward mono camera, forward-facing thermal and visible wavelength cameras, a radio modem to communicate between vehicles, and a Li-ion smart battery. The individual components of Gen 3 have been tested both in simulation and and real life, and they show promising results.

The final simulation paper by Lee et al. discussed an Adaptive Stress Testing (AST) framework to find which path forward is most likely to cause a failure in simulation [56]. The simulation was generated using a Monte Carlo tree search algorithm. They use a hidden Markov model with Reinforcement Learning to find the path. The results show a safer alternative to what was currently used.

4.1.4. Other

This subsection highlights aviation systems that do not fit into one category. In 2019, Gilbert et al. released a paper which goes over a flight test program for testing computational resources used in AI and ML applications [92]. Since testing in space is very expensive, high-altitude balloons offer a cheaper alternative that can provide a facsimile of the space environment. Communication in space requires high-gain equipment that is expensive to maintain and hence has no assurance of being in the budget for a spacecraft’s entire mission. Because of this, on-board data processing is preferred. High-altitude balloons are simple and cost effective, only being about USD 1000 a pop (pun intended). This communication from space to earth is maximized for communication strength with a neural network, which requires dedicated acceleration hardware [92].

Dolph et al. presented an image-based aircraft detection and tracking pipeline for motocopters, fixed-wing, and normal-sized planes [93]. They tracked two aircraft at once, always tracking one motocopter. The pipeline used both image differencing and morphological detection with a Kalman filter-based tracker. The overall method returned a good track rate of 71% [93]. In 2021, the research team released another paper about tracking aircraft [94]. This system detects and tracks aircraft with only optical sensing. Multirotor and fixed-wing sUAS, as well as general aviation planes, were used for the study. The methods developed in this study achieved 90% accuracy in flat terrain but did not perform well in terrain with roads, rivers, mountains, etc.

In 2020, NASA released a paper which covers an ML-enhanced engine design for predicting cruise TSFC (a measure of fuel efficiency) [57]. Predicting this measure can help greatly increase the engine design time. The algorithm achieved a TSFC prediction accuracy of 98.3% and perfect core size prediction, allowing it to be used to predict which engine will perform the best among several candidates [57]. In 2023, Tong released another AI-enhanced airplane engine design that saves time, reporting promising preliminary results [63].

4.1.5. Open Challenges of Aviation

The airspace category has its own unique capabilities and limitations, as outlined in

Table 4. This table shows the capabilities and limitations of Section 4.1.

Capabilities	Limitations
AI has been used for ATC systems.	Safety assurances for AI systems need to be made.
AI systems perform well at aiding in navigation	Human-AI interaction needs to be improved
AI systems have proven capable of aiding in system control.	Curated datasets need to be publicly available.

at the beginning of this section. The most promising open challenges have to do with natural language processing, AI-enhanced ATC, and safety guarantees for AI systems.

AI has many ways of interacting with human operators. Over the last decade, research into AI with natural language processing has proven to be a viable method for pilots and other users to interact with AI. Further study needs to be undertaken to (1) improve the performance and (2) to measure if the user experience is improved using language-based communication rather than traditional written methods.

While AI-enhanced ATC systems have been gaining traction over the past two decades, additional progress can still be made. More efficient algorithms (both in speed and performance), more capabilities, and more assurances are all possible avenues to pursue. To this end, creating well-balanced datasets is important. Anomalies represent a relatively low percentage of ATC; if anomalies have low representation in the training sets, then the AI normally has low performance in identifying anomalies, reducing the overall safety.

Safety is a concern in aviation systems, and it is perhaps the largest future direction identified in the surveyed literature. Safety assurances are harder to show than good performance, because the system ought to be safe in all circumstances. Because of its complexity, we have devoted Section 6 to a full discussion of certification, verification, and validation. The aim of assuring these systems is to create reliable systems that the end user feels comfortable enough to use. However, even if the AI has been shown to be reliable, the end user might not adopt the system. Thus, another question for further research is “What convinces an end user to adopt safety-critical AI systems?”

4.2. Space

Bandwidth and latency issues make it infeasible for earth-based computers to command and control distant spacecraft. AI is thought to be one solution to this problem [5,95]. AI-based assistance can come in the form of anomaly detection, schedulers, AI crews, medical support, and communication. Unfortunately, running AI models beyond Low Earth Orbit (LEO) can be difficult because high-performance computers are heavy (i.e., costly to send to space) and can behave unexpectedly when not protected by Earth’s magnetic field [96,97]. Certain capabilities, such as repairs and docking, require robotic systems to be designed to handle the fragile and expensive components [10]. In addition to these physical concerns, AI systems need to be robust and consistent enough to be trusted with the lives of the crew [98].

In this section, we will examine the capabilities and limitations of NASA AI systems in space from 2008 to 2024. Through this examination, we hope to gain insight into what these AI systems are capable of handling and in turn how much to trust these systems.

4.2.1. Anomaly Detection

Crews are becoming smaller and hence are unlikely to possess all areas of expertise necessary to detect and mend faults. As such, the need for autonomous anomaly detection becomes paramount. Anomaly detection can alert the crew when system failures occur, allowing a problem to be resolved more quickly than would otherwise be possible.

A 2008 report on the Inductive Monitoring System (IMS) deployed on the International Space Station (ISS) describes how the IMS can determine how much the ISS is deviating from normal sensor readings without the need to have previously seen the anomaly [99]. The IMS reports all faults and does not give indication as to where the fault originates, and so it requires an experienced crew or ground support to diagnose the problem. By 2014, the IMS had been used for the Advanced Caution and Warning System (ACAWS) which, in addition to fault detection, also performs automated diagnostics to find the root cause of the problem [100]. ACAWS is a generalized system that is able to encode 80–90% of failures without being designed with a specific system in mind [100]. This flexibility and root cause detection allows for a more generalized crew.

In 2018, an automated power system was created with fault detection for power controller faults, line faults, shorts, sensor faults, sensor bias, loss of signal, and undue noise [101]. Using this automatic controller places the power system in safe mode after the fault is detected. The controller also attempts corrective actions for planning and system operations, though it is unable to correct hardware problems.

Fault detection has grown substantially in the past 15 years, now allowing for near-autonomous detection and diagnosis. Anomaly detection can enable “self-healing” systems, meaning that through fault detection and identification, fault healing can be accomplished with only minor physical intervention [102]. This vision has yet to be fully realized and thus it is important to remain vigilant while using anomaly detection software.

4.2.2. Scheduler

System-level plans are required to ensure that a spacecraft can function. These plans have historically been made entirely by ground controllers [103]. Ground-based solutions become infeasible as space exploration moves beyond LEO missions and into missions where communication delay becomes significant [101]. In 2006–2012, the unmanned flight groups began communicating with the manned flight centers to share their autonomous technology that was previously compartmentalized [104]. These automated systems included schedulers for mission planning up to two weeks in advance, taking some of the burden off of the ground team.

As early as 2014, NASA created an architecture for an autonomous power control capable of collecting power information and using that information to maximize a vehicle’s resources to complete mission objectives [103]. Later systems also included the ability to detect faults and safeguard electronics against in the event that a fault was detected [101]. The automated system also allowed for a new plan to be made if a fortuitous event should occur, resulting in more efficient systems.

Power is not the only thing that needs to be scheduled. In 2015, an AI capable of planning the mission, engineering, orbit allocation, and instrument use was created [44]. This allows for more efficient use of the crew’s time for scientific endeavors. A similar AI model was used to efficiently plan the Mars Perseverance rover [105]. This planner is able to use the limited computational power found on the rover to schedule tasks and take advantage of using fewer (or more) resources to plan the rest of the day [105].

The need for automated schedulers that can adapt to mission-specific needs continues to grow [98]. Schedulers have the capacity and ability to streamline the planning process. Mission-specific schedulers have performed well. However, as they become more common, it is important to know the bounds of their capabilities in order to judge how much trust ought to be put into them.

4.2.3. AI Crew

Some AI systems are designed to be more general purpose. We have labeled these systems “AI crews” because of their added flexibility and breadth. These can take many forms, such as a collection of algorithms compiled in one place for ease of use, real language models, or speech recognition [106]. The NASA Autonomous Systems and Operations (ASO) division has developed AI capable of completing the tasks of three different flight controllers: power, thermal, and life support [107]. One of these is the ACAWS power monitoring system explained earlier with the additional ability for the user to create hypothetical “what if” scenarios [100]. This allows for the user to prepare for outcomes that have not physically occurred before.

Life support systems are necessary for space travel; long missions assume that nearly closed oxygen and water cycles will be achieved [108]. This can be accomplished with an AI system that controls the oxygen and water loops in a spacecraft [108]. Life support has yet to be proven to operate as a closed system using AI without ground input, so human missions to deep space are not yet feasible [108].

In 2020, a Lunar Autonomous Position navigation and time System (LAPS) was designed in order use a small swarm to gain position, navigation, and time information [109]. This system would be smaller and easier to manage compared to the Global Positioning System (GPS), and it has a simulated performance of 10 m [109].

NASA has also considered the possibility of using speech recognition technologies to allow for human–machine communication [75]. The space station can be noisy, and the microgravity in space can make the human voice sound different [75]. These challenges can be accounted for through careful consideration of microphone type, noise suppression, and vocabulary selection [75]. By considering these principles, NASA achieved a 97% accuracy after retraining a ground-trained speech processing model on the astronauts in space [110]. In order to develop trust in speech recognition in space, it must have a high accuracy and be robust enough to work as expected in unusual circumstances. If speech recognition is to be the primary form of human–machine interaction, it must be easier and more convenient than typing or another form of a controller. The high accuracy shown in [110] demonstrates the potential benefit that speech recognition can serve.

AI is also advantageous to have on unmanned missions. AI can “crew” unmanned deep space missions and even process data on site [111]. This is necessary because as a spacecraft increases its distance from Earth, the communication bandwidth decreases. This makes it unfeasible for information to be processed on Earth, and so it must be processed on location [111].

In the presence of a human crew, AI crews tend to serve as aids to make the work lighter, but they do not completely alleviate the burden of work from their human counterparts. This is partly due to physical limitations, such as an AI not being able to fix a hardware problem, and partly due to constraints in the capabilities of these AI systems.

4.2.4. Medical

Medical problems can arise during space exploration, and communication with ground-based doctors may become unfeasible as mission distances increase. Because of this, NASA has looked into many AI solutions to arising health problems. These AI solutions come with their own pitfalls, such as a limited amount of astronaut training data, as well as ethical and legal concerns associated with using an autonomous system for human health [98,112].

Due to the difficulties and expenses associated with developing models in space, most of the medical AI systems developed are first tested on the ground. In 2013, NASA analyzed posture to check if astronauts were appropriately adjusting to earth gravity after trips to space [113]. This same algorithm also has the capability to help physical therapists see how well patients are recovering after a stroke [113]. NASA later developed an ultrasound system that allows for automated medical advice that the chief medical officer can use to inform medical diagnostics [114].

In space, NASA successfully utilized a CNN-based computer vision model to non-invasively gain posture information of astronauts while wearing space suits [115]. This posture information can be used to characterize the risk of injury associated with the task the astronauts are assigned [115].

One important medical application relating to space has to do with radiation. Beyond Earth’s magnetic field, an astronaut can expect to be exposed to over 15 cGy of individual Galactic Cosmic Radiation (GCR) ions [58]. Scientists at NASA have run experiments on rats to predict if there would be any cognitive impairments after being exposed to GCR [58]. The results were processed using an ML algorithm capable of achieving an accuracy over 50%, which is better than a random guess [58].

Looking to the future of Earth Independent Medical Operations (EIMO), EIMO needs to be able to do pre-mission planning, acute and prolonged medical decision making, supply and resource management, and task load balance with a language overlay that allows for ease of use for the crew [98]. At a technical interchange meeting in 2023, NASA scientists decided that the best way to achieve the goals of EIMO was to have several distinct AI models that each serve a more specific purpose. These models will then be unified into one overarching model [98].

With medical AI, human safety and legal concerns require more caution. There are not yet any models that directly diagnose problems; existing models only aid in the diagnostic process. It seems possible that medical AI could serve future manned missions; however, it is currently unable to completely fill that role.

4.2.5. Communication

In 2013, NASA released a report on cognitive network solutions for more effective communication using AI and ML [116]. This enables networks to detect and repair anomalies, increasing performance reliability and security while reducing costs. By 2019, NASA had envisioned a cognitive networking solution called a “space cloud” that allows for distributed cognition, decentralized routing, and on-orbit data processing [102]. These improvements are anticipated to be able to increase the mission science return.

In 2020, NASA began developing a cognitive network that would learn if something scientifically interesting was going on and then divert more resources to capturing the event leading to better information on the phenomena [117]. Sometimes, continuous observation is required to properly monitor a phenomena. The Multi-Agent Clusters for Persistent Observations from Space (MACPOS) project at NASA is researching a system of individual agents that can autonomously create a formation that continuously covers a spot [9]. As of 2021, they had began testing on ground-based systems, but the algorithms were still under development.

Over the past decade, AI has shown the capability to improve communication, resulting in better scientific outcomes. This field appears to be growing, and so it is probable that the success of AI in improving communication efficiency will continue.

4.2.6. Satellite

Satellites can be used as effective remote sensing tools, as they quickly and effectively cover large areas of the Earth’s surface. Remote sensing data can be used for many applications, including weather detection, public health, and agriculture.

In 2019, NASA used an ANN on data from the Aqua, CloudSat and CALIPSO satellites to detect multi-layer ice-over-water clouds [55]. Detecting clouds can give us insight into the atmospheric energy system. The algorithm created in [55] was able to achieve a better total detection rate than previous methods, reporting a combined single- and multi-layer detection rate of up to 80% in the day and 77% at night.

Unfortunately, satellites cannot always cover the entire desired area. For example, satellite tropospheric N${\mathrm{O}}_2$ from the TROPOMI and OMI satellites only covers about 51% and 87% of China, respectively [60]. N${\mathrm{O}}_2$ is a threat to both the public and the environment, and so this NASA-funded study produced a combined missing extra-trees and deep forest model to fill in the gaps that are present in the satellite data. This endeavor was successful with average R-squared values ranging from 0.89 to 0.96.

As climate change continues to make more volatile weather, surface temperature readings with a high spatiotemporal resolution become more important [61]. Since a single satellite cannot offer this high resolution, NASA used an AI model to combine many different satellites with differing orbit patterns to achieve a mean absolute error of 1.73 K and a resolution of 2 km every 10 min, which is much higher than the standard spatiotemporal resolution of 10 km every hour. Climate change also has an effect on the water cycle, and it may have an effect on aquifers around the world [64]. Since an estimated 43% of irrigation water comes from an underground source, it is important to monitor the health of aquifers. Some of these aquifers do not posses the equipment to monitor the spatiotemporal dynamics of these underground water sources, and so NASA turned to using satellite data to monitor the aquifers present in Brazil, which is a nation that had been in a prolonged drought at the time of the study [64]. They were able to measure the aquifer groundwater storage change using gradient decent, linear regression, and wavelet and seasonal decomposition applied to satellite gravimetric data. They achieved an RMSE of 2.68 cm, outperforming the Global Land Data Assimilation System.

Enhancing satellite-based remote sensing with ML has provided good results. ML is well suited for both regression and classification problems, which are often the subject of satellite-based remote sensing.

4.2.7. Open Challenges of Space

AI applications for space are varied. The capabilities and limitations are summarized in

Table 5. This table shows the capabilities and limitations of Section 4.2.

Capabilities	Limitations
AI has proven capable of being an adaptive scheduler.	Assurances of performance in possible scenarios are needed for AI systems.
AI systems designed for anomaly detection perform well at their function.	“Self-healing” systems are thought to be possible through AI.
AI systems perform well on satellite data.

. The most direct avenue of future research is to target existing satellite data with AI and ML techniques for new applications. Important climate data have been shown to be retrievable through applying ML to satellite data. Thus, it stands to reason that further climate data could be retrievable through applying ML techniques to satellite data.

Autonomous anomaly detection was made with the intention of building self-healing systems that have yet to be fully realized. Self healing is the ability of a system to repair faults by itself through autonomous fault detection and identification with only minor physical intervention. Current systems lack this ability, so future research could fill this gap to create self-healing systems. The application space of the system is less important than to show repeatable steps for how to make a viable, reliable self-healing system. Further research into how to make self-healing systems in specific domains will also be important, because different challenges will arise in different domains.

Space communication has proven to be one such area where self-healing systems have shown success [116]. However, further research needs to be conducted into different applications of AI-aided communication.

Additionally, because of communication delays related to space travel, adaptive schedulers are important. While much research has been conducted into creating adaptive schedulers, increasing the performance and reliability of these systems remains an open research objective.

In the field of AI for space, assurances for AI systems remain to be an open challenge, being identified directly for autonomous crew and medical AI systems. Because of the complexity and importance surrounding the verification of AI, we have devoted Section 6 to a full discussion of certification, verification and validation.

4.3. Platforms for Automation

Platforms for automation can take many forms. The broad definition of platforms that we use is that it is a process or project that is easily applied to different missions.

It is sometimes difficult to know how easily one project can be modified such that it can be used in another project. For example, can a power analysis tool such as [101,103] be used or adapted for other systems? Because of the difficulty of accurately defining what tools can be considered platforms, in this section, we only discuss literature that expressly states to have been used on multiple projects or built with modularity in mind. Platforms are necessary because as space travel ventures past LEO, the communication delay becomes too great for humans to operate the systems from earth in real time [100].

4.3.1. ATTRACTOR

Autonomous systems are required to be both proven and perceived to be capable [118]. AI systems learn through experience, so they often do not generalize well to new environments [7]. Simulations grant a platform for growing past this hurdle, allowing for AI to experience new things in a safe and easily modular environment [7]. NASA’s Autonomy Teaming and TRAjectories for Complex Trusted Operational Reliability (ATTRACTOR) project aims to provide a platform for understanding when an autonomous cyber–physical–human system is trustworthy and trusted. [118,119]. Trustworthiness is related to the proof of system capability, and trust is related to the perception of system capability [118,119]. Many ML algorithms do not provide explanations for their answers and thus have low explainability and trust [118]. Unity—a popular game engine—can be used to provide a lens through which explainability, trust and trustworthiness can be examined [118].

Another platform for improving trust looks at explaining complex neural networks, including CNNs. A CNN with many layers is unexplainable. However, the individual layers can be examined to gain insight into what features are being represented [13]. The ATTRACTOR group examined what information could be extracted from these layers and determined that the first few layers hold information about edges, corners, and spots, and further layers are more difficult to interpret because they are highly dependent on nearly all of the original pixels. For the deeper layers, a deconvolutional approach can be used to interpret the layers. Improving how well these models are understood can lead to improvements in trust, quick retraining, and an improved design of CNNs.

The ATTRACTOR project then turned its focus to creating a simulation of an AI search-and-rescue operation. First, they examined how questions about the missing person can be posed in order to ensure accurate descriptions of the person are obtained [120]. They used two different prompts (one aimed at a general description and another with the express purpose of aiding in the person’s rescue) and two different mediums (text and speech) [120]. The study found that when asked to aid in the rescue of a missing person, the participants described colors more often and activities less often. The study also trained ML algorithms to detect when people were not describing the scene well. NASA then looked more closely at the descriptions, training several ML algorithms to predict context from the image description [121]. They determined that asking to describe a missing person through text was the best for search-and-rescue target description, and they made an ML algorithm to provide feedback on if the user’s response was good enough.

The ATTRACTOR team then developed a simulated multi-Unmanned Aerial System (mUAS) with fault detection in order to search for the previously described missing person [89]. It is not easy to determine when a fault occurs because many anomalies are previously unknown, so the researchers decided to train a model to represent normal behavior to see how well the current behavior matches normal behavior. This is accomplished using an active learning method that allows for less time spent by experts in the field to improve the detection algorithms. Furthermore, the AI system is capable of predicting when an anomaly is going to occur. Several faults were simulated in Unity; the flexibility provided by Unity allows for more generalization.

The ATTRACTOR platform allows for a glimpse into human–machine and machine–machine interaction [122]. This examination led to insights on the trust and trustworthiness of the AI systems as well as to a reflection on how to connect the AI systems. This is valuable information to have as more and more autonomous systems are used in conjunction with one another and with humans.

4.3.2. Others

In 2008, many of the computer helpers were called “expert systems.” These expert systems can be categorized as a computer program that draws from the knowledge and analytical skill of human experts to provide a digital effigy of them. One such example of an expert system is the Spacecraft Health Inference Engine (SHINE), which is designed to aid a human operative in monitoring, analyzing and diagnosing the health of space flight systems [12]. SHINE was purposely designed to be easily to embedded into other systems. Because of this ease, SHINE was used in over a dozen NASA and commercial projects [12].

By the mid 2010s, NASA combined the ACAWS SPIFe and EUROPA into one platform with an improved GUI [100]. This platform allows for a smaller crew of astronauts to pilot a spacecraft because the combined platform can aid in many different tasks, allowing for a trained non-specialist to complete more specialized tasks.

A different NASA-funded project integrated health management systems with AI to make an autonomous system capable of increasing both safety and reliability. The techniques shown in Medina et al. are meant to be an example of how to develop autonomous control systems quickly [123].

Large datasets are not always feasible to obtain, and as such, ML with a quicker learning curve is required. Wong et al. discussed a knowledge bot-based deep network that uses Reinforcement Learning and manifold learning with an SVM output layer to reduce the amount of data required to train [50]. Knowledge bots are used to bypass running simulations, take input parameters, and quickly tell the results as if the simulation had been run. The training on this knowledge bot was able to be reduced from the previous 10,000 training simulations to 25. This will greatly reduce the amount of simulations required to train future ML algorithms.

By the late 2010s, NASA had developed a different biology-informed approach for signal processing [124].

In 2018, NASA used IBM’s Watson to make PEER. PEER is an AI advisor that provides relevant information for real-time decision aiding, and it is supposed to put the information where it is most needed for quick understanding [52]. The full system was not developed at that time. However, a prototype was made that allows users to ask questions and receive accurate answers [52].

From 2016 to 2020, NASA worked on the NASA Platform for Autonomous Systems (NPAS) [125]. This platform is supposed to enable AI to evolve past traditional single-use-case algorithms into a more innovative general purpose AI. To achieve this goal, NASA plans on taking advantage of redundancy and repeating commands, integrating System Health Management to detect anomalies, create object libraries, gain knowledge of model applications, and perform AI scheduling. NPAS boasts being applicable to many different systems, as it is built with modularity in mind, lower cost, and quicker. NPAS also enables autonomous applications to talk with each other.

In 2020, NASA published about LAPS, which is a platform developed with configuration in mind [109]. It is capable of simulating a multitude of scenarios. While the platform is not at the forefront of the paper, it is still quite important and shows the value of simulations for AI development and testing.

A different paper outlines a general approach for quantifying uncertainty that would need to be tailored to each individual system [126]. As autonomy continues to increase, knowing the capabilities and conditions of a system is required to know if it can be trusted to be deployed. To understand this, the sensor performance, degradation, and system behavior need to be understood. These are inputs to the system that will more holistically quantify the uncertainty of the system. At the time of the paper, this system did not include room for modeling natural unrelated uncertainties, such as wind, but there were plans to add this capability.

ML is a powerful tool that can be difficult to use when the user does not fully understand how the models work and when to apply them [59,127]. In order to combat this, NASA developed an ML tool that includes 44 different regression algorithms and many pre- and post-processing methods. This enables a user to create a ML pipeline, or it can be automated for them if they are not as familiar with ML techniques. This tool allows for the less knowledgeable to try out many different models easily, meaning the lay person can utilize ML with success.

Tong outlines Aero-Engines AI, which is an ML tool to asses aircraft engine designs that is user friendly [63]. This tool uses a KNN deep learning approach and has a GUI for ease of use. The paper claims it saves a significant amount of time in design exploration. It also claims it would be easy to expand the types of engines it can operate on.

4.3.3. Open Challenges of Platforms for Automation

The main capabilities and limitations for platforms for automation can be seen in

Table 6. This table shows the capabilities and limitations of Section 4.3.

Capabilities	Limitations
Platforms for ease of AI application have been made, and work well.	Simulations have yet to be proven to be a viable way to prove performance in the real world.
AI platforms have been designed for multiple uses successfully.	Human–AI interaction has been studied. However, it can still be a limitation for use.

. Because platforms for automation focus on usability, they allow us to see many open challenges in human–AI interaction. The main aspects of the interaction examined have been explainability, user trust, and usefulness. Explainability is the ability of the end user to understand the algorithm used. For complicated models, such as an ANN, this would mean creating methods of discovering why the model is predicting as it does. More research into explaining complicated AI models would allow for developers to gain the higher performance benefits of current black-box models while not sacrificing explainability.

User trust can be a complicated topic, but it is necessary to understand if we wish for users to be able to rely upon AI systems. However, user trust is subject to change, so both obtaining and retaining user trust will remain a constant open challenge.

As was noted in previous sections, performance assurance remains a concern. Simulation space has been identified as a possible path for assurances. However, the simulations themselves must be assured as well, which in itself adds another open challenge. Simulations are meant to mimic the real world, and so they must mimic the uncertainties of sensor performance. Thus, the challenges of both measuring the uncertainty and simulating the uncertainty of sensors are potential future research directions. Because of the complexity and importance surrounding the verification of AI as a whole, we have devoted Section 6 to a full discussion of certification, verification and validation.

5. Synthesis

AI at NASA is varied in both application space and algorithms. These AI systems are designed with the purpose to be used, and so they must both be proven and perceived to be capable. In order to perform well, popular AI algorithms, such as neural networks, need large amounts of balanced data. In many cases, large and balanced datasets can be difficult to acquire. For example, in aviation datasets, anomalies are rare, but they are very important [89]. The imbalance of the dataset could lead to misidentification of the anomalies, which poses a safety risk. In some fields, such as protein modeling, it is impractical to acquire large datasets [67]. When large datasets are unavailable, ANNs do not perform as well as more traditional methods such as SVMs; thus, it is important to consider the amount and balance of the dataset when choosing the AI algorithm.

For AI systems to be perceived as capable, the models need to become more interpretable. This has primarily been accomplished through gaining insights into previously black-box models [13] or using PBM [79]. Both avenues offer advantages and disadvantages. PBM chooses or builds the model in tandem with a known process, so it can leverage the advantage of representing a real system with known mathematical relationships. This can create a very interpretable model. However, not all systems have known relationships. In this case, it is better to gain insight into a more general model. These insights can tell us why the algorithm predicts as it does.

With any AI system that interacts with a human, assurances are required to show safety and repeatability. Unfortunately, systems that involve human–AI interactions go past interpretability and are not fully constrained, making them difficult to design. A major challenge of AI systems are that they are non-deterministic, and they can have different outputs even when the input data are similar (or identical). This creates a major adoption limitation. We have devoted Section 6 to discussing the challenges surrounding the certification of AI systems.

6. Adoption Limitation—Certifications, Verification, and Validation

As mentioned previously, one large open challenge is that of certification, verification, and validation for autonomous systems, which represents a large adoption limitation. In reviewing the literature, most papers focused on the limitations of adopting AI systems rather than the limitations in the capabilities of AI systems. Because of the attention certification and verification received throughout the reviewed literature, we dedicate this section to exploring three publications that focus on these topics [128,129,130].

Before diving into these three papers, we present some important related findings that were discussed in the Workshop on Assurance for Autonomous Systems for Aviation [7]. The conference notes that while AI can grant many benefits over human operators, its behavior can be irascible and difficult to predict. Nearly every part of the National Airspace System is becoming increasingly autonomous, and these autonomous systems need to be verified in order to be implemented safely. When evaluating the models, it is important to look at what it must do rather than evaluate the outer reaches of its capabilities. The Workshop on Assurance for Autonomous Systems for Aviation [7] showed a preference toward probabilistic reasoning techniques for verification. This is because while AI can have non-deterministic behaviors that are hard to pin down, probabilistic reasoning can measure AI certainty for a particular course of action. The workshop also proposed a change in the certification of AI. Participants suggested changing a “certification” (i.e., a declaration that an autonomous car is safe) could be changed into a “license” (i.e., granting the AI permission to drive a car) [7]. This comes with complications. For example, in the scenario of a self-driving car that has human input, how can a license be provided when it also relies on the human? While these questions do not require an answer to produce the AI, they are required for certification, verification and, in some cases, use.

6.1. Autonomous Systems as Decision Aids

AI is only as useful as its ability to synthesize information and present that information in an easily digestible way. Mosier et al. gives examples of how AI for plane system management can be confusing to interpret, causing wrong decisions to be made in life-threatening situations [128]. In order to effectively aid decision making, the AI must act like another crew member or risk being harmful because of the increase in superfluous information. The paper offers two different options: decision aids, which provide information on the status; or specified mitigation actions, which tell the pilot what to do. The decision aids tended to have more of an impact; naturally, the impact was positive if the AI provided correct information and negative if the AI provided incorrect information. Since the AI can have a large impact on performance, the question then becomes how much we ought to rely on this tool.

Table 7. This table shows the different levels of autonomous systems according to [128].

Level	Description
1	Human makes all decisions
2	Computer generates options that the human can ignore or use
3	Computer selects generated option, but the human can choose to do it or not
4	Computer selects best option, human can still change option
5	Computer controls humans can intervene
6	Computer controls everything

shows the different automation levels that can be used to measure reliance on automation [128].

It is also important to keep in mind the “automation paradox.” When using AI, the user’s skills atrophy—either by the user not gaining skill or by eroding a previously held skill [52]. Because of this, an AI that only requires human input in dangerous situations might be more detrimental than having an AI assistant that provides helpful information. The change in human input could also lead to changes in how we certify these systems.

When considering how much trust to put into AI systems, it is also important to note that humans have bias. For example, when a user is tasked to monitor the performance of AI, the user’s trust in AI tends to decrease [128]. Research has also identified a strong correlation between trust and reliability [128]. Thus, in order for AI systems to gain and maintain trust, they must perform as expected—especially when being actively observed by a human [128].

6.2. Safety Concerns and When to Use Adaptive Algorithms

AI has many different challenges associated with certification compared to standard algorithms. The following is a summary of the discussion found in [1].

Systems ought to perform as intended under any foreseeable operating conditions. However, additional restrictions must be considered when dealing with safety-critical operations. (For example, in an aircraft, a failure that would prevent landing is required to be much less probable than the failure of an indicator light.) For software, the assurances include not having any unintended functionality. This is difficult to show with adaptive and stochastic systems, so the newest version allows the use of analysis and other techniques to meet this requirement. The verification process has two steps: (1) check if requirements are met and (2) verify a high confidence that all errors leading to failures have been removed.

A large certification challenge is that adaptive systems change their behavior over time. Designing tests to assure that there are no unintended behaviors is difficult. The code must also be well documented. In addition, certification required full control of the source code (which requires freezing open source libraries that are subject to change), traceability, and a transparent design. Finally, code must be written in a conventional way with low complexity. With these restrictions and requirements, adaptive systems are more difficult to make and validate than standard algorithms. To determine if the benefits of an adaptive system outweight the costs, the following questions were proposed:

• Does the adaptive system perform significantly better?
• Is the algorithm reliable?
• Is there an argument for why it is reliable?

If the adaptive system does not provide a significant improvement to the standard system, or if it is not apparent that the system is reliable, then an adaptive system is likely not the best solution.

In 2015, public support for AI was low, so certification and validation techniques for adaptive systems did not exist. Regulators have an expertise gap with AI, and the aviation industry is hesitant to adopt technology without thorough testing. Education helps overcome the perception barrier, and it will help people utilize the tool of AI. In order to make AI certification a reality, the following suggestions were made:

• Relax the static behavior requirement within certification standards.
• Make certification and verification methods more flexible.
• Adaptive systems could obtain licenses similarly to how humans obtain licenses to perform specific tasks.

An example for a possible verification technique is found in the study of Pasareanu et al. that was released in 2023 [129]. DNNs for perception are difficult to test because of their size, lack of specification, and sensitivity to environment. Current verification techniques can only work on modestly sized networks, and they are unable to cover the entire input space. Pasareanu et al. [129] suggest a compositional verification approach, which takes both conventional components and learning-enabled components to test a desired behavior. The system is analyzed without the DNN to determine what behaviors of the DNN are required for safety. Then, assumptions about the DNN are made to determine the output requirements of the DNN. The output of these assumptions can be used to train and test DNNs. Experimental results on an autonomous taxiing system for aircraft show that the mathematical assumptions necessary for training and testing are small and understandable even though DNNs have large output spaces. Probabilistic analysis is a good tool for estimating if the extracted assumptions are violated in real time. While this technique shows promise, it can suffer from scalability issues. Fixing scalability is an ongoing effort.

6.3. Certification in Aviation

In 2023, NASA scientists convened with industry and institutions of higher education to create a roadmap for the future of verification and validation techniques for autonomous systems [130]. The following is a summary of their discussion and resulting paper.

The next generation of aircraft will have more AI than the current generation, and these AI tools should offer benefits without decreasing safety. AI and ML are adept at tasks involving perception, localization, mapping, planning and control—all of which need to be verified. Previous verification and validation techniques do not fit well with data-driven AI, since AI can behave unpredictably. Older techniques focus on behavior rather than data, and so it is difficult to have a one-to-one technique that matches previous verification and validation techniques with current AI systems.

The airspace industry continues to grow, and there is a shortage of pilots. AI may be able to help alleviate the shortage. Some needed advancements in AI systems in aviation are to improve AI capability to achieve the following:

• Limit the unexpected behavior from AI;
• Have good human–machine interaction, including the ability of an algorithm to detect if the pilots are alert and take over if they are incapacitated;
• Account for new vehicles and AI in airspace;
• Use computer vision to land;
• Taxi and avoid objects autonomously;
• Perform onboard flight path finding; and
• Improve safety over time.

There are many safety assurance concerns when using AI in a safety-critical area such as aviation. In the short term, there is a need to certify that these AI provide stability and meet expectations. Part of this need will be addressed by changing existing tests to be better suited for AI. This could include having run-time assurance that can be used to stop unwanted behavior.

The ability of AI to have control over emergent properties is called compositionality [130]. Reinforcement Learning is a promising way to achieve compositionality, because the developer can either reward or punish behaviors. There is difficulty associated with proving the absence of emergent behaviors in order to ensure the safety of the system. Simulations can identify emergent behaviors, but they cannot prove their absence. In addition, there is often a significant gap between simulation and reality. The difficulty associated with ensuring the safety of an algorithm makes it difficult to certify and trust the systems, and there is no standardized way of accomplishing this. The current best practice is to use statistical testing to verify ML algorithms. These statistical tests aim to verify correctness, robustness, and a lack of overfitting. In the short term, ML algorithms need assurance techniques; in the medium term, the techniques developed need to be standardized and adopted; and in the long term, trustworthiness needs to be quantified.

There are three primary levels to human autonomy interaction: coexistence (when autonomous system and people work in the same space), cooperation (when autonomous system and people complete work toward the same goal in the same space), and collaboration (when autonomous system and people work on the same task together) [130]. Human–AI interactions introduce both physical and psychological safety concerns. Physical or psychological stress or discomfort introduced by an autonomous system can have serious negative health effects. Psychological harm can come from robots breaking social norms. In addition, physical harm can cause psychological harm.

Run-Time Assurances (RTAs) add assurance components that constrain complex behaviors so there are no unsafe or unintended behaviors. RTAs can include the following:

• Run-Time Monitors —Run-time monitors continuously check variables related to the system state; if the input or output is outside of normal bounds, the monitor can stop or replace the system.
• Safety Backup Components—Safety backup components provide alternative functionality (already verified to be safe) that can step in if ML behaves unexpectedly.
• Switches—Switches manage whether the complex component (AI) or a safety backup is being used.

These RTAs have shown feasibility in Vehicle Management Systems (VMSs). The four run-time monitors created for this system preformed as expected and kept an aircraft safe [130]. VMSs can also be used for urgent landing planning.

The roadmap [130] proposes a similar validation process as is found in the pharmaceutical industry. In this framework, an autonomous system would be approved for limited use in contexts with oversight. Use in other contexts would be granted as the system proves its ability to safely and effectively perform tasks. Dynamic assurances, such as what is found in the pharmaceutical industry, allow for ambiguities. Dynamic assurances are when specifications and assurances are changed in real time to better match the real-world requirements. This can be achieved by intertwining simulation and physical testing. These changing assurances exist in “stages” defined in terms of specific sociotechnical contexts. The stages are (1) early-phase testing that verifies the integration and implementation and develops appropriate base models for offline testing, (2) transitional testing that identifies the context of failures in real life and establishes how the environment might change, and (3) confirmatory testing that determines system performance and reliability and the extent of the system–user value matching. Implementing a dynamic assurance technique requires advances in co-design, simulations, testing, and legal practices for deployment stages.

New applications for autonomous systems include aerial inspections, package delivery, first responders and advanced air mobility [130]. The functions required to accomplish this are flight planning, performance authorization, flight approvals and detect-and-avoid (DAA) functions. UAS Traffic Management is supposed to prevent collisions by using Strategic Separation (pre-flight), Tactical Separation (during flight—longer time scale) and Onboard Collision Avoidance (during flight—short time scale) that are similar to how current Air Traffic Management systems work.

Certifications for software are difficult. Difficulties in the certification process include a need for better verification techniques, legal concerns, and difficulty in changing software. Additionally, current certification processes suffer from a lack of common requirements, quality data, understanding of human–machine interaction, trust, and assurances. Another challenge is the wide variety of functions that software can accomplish. Software can be used for perception, mapping, planning, control, coordination, and subsystem management. These tasks can take different forms, such as autonomous takeoff and landing, autopilot, collision avoidance, system-to-system interaction, and system-to-human interaction.

When looking at AI and ML systems, the challenges of assurance further increase because of the difficulty in proving no harm will take place through unintended behavior. Additionally, in ML and AI software, it is difficult to determine which line of code is responsible for which behavior, making it even more difficult to prune unexpected behavior. Formal analysis and structural coverage analysis can be used to determine that there are no inputs to the system that result in an observable unintended behavior [130]. If unintended actions are present in AI or ML, they can be mitigated by having a complete training and test dataset, better structural coverage metrics, and run-time assurances. ML techniques were not initially designed to provide guarantees of correctness.

Despite all these challenges, it is suspected that it will be easier to verify and validate low-impact and low-complexity AI models [130].

7. Future Directions

There are many applications for AI systems and hence many directions for possible future research. However, it is outside of the scope of this review paper to recommend future directions for each individual AI application, and we will instead focus on long-term goals that apply to multiple applications. AI has greatly increased in capabilities in the past three decades and as such has seen a wider adoption for scientific use. Improving algorithm performance and identifying novel applications are continual sources for future research.

AI must be verified and validated prior to being adopted into applications such as aviation, space travel, or other safety-critical areas. Current techniques for verification and validation have been lagging behind the progress of AI algorithms. Therefore, verification and validation techniques are a source for improvement. These techniques can include understanding why AI systems behave as they do, creating metrics to statistically show that the system is highly likely to behave as expected during its use case, or performing a formal analysis to determine that there are no inputs to the system that result in unintended behavior. Simulations have been shown to be a promising way forward for determining if an algorithm behaves as desired. In addition, Reinforcement Learning is thought to be a good path forward for ensuring an AI behaves as expected.

Recommendations for Future Reviews

In addition to the future direction of improving AI systems, we note that there is work to do in synthesizing the literature in the field of AI. For example, our paper collection methodology (described in Section 1.2) focused on analyzing papers in the NASA Technical Reports Server using specific search terms and internal relevance sorting. A future review could employ a more principled approach, such as the PRISMA method [131].

As noted previously, the papers resulting from our literature review search focused their analysis of limitations on current challenges that limit the adoption of AI into NASA systems. Thus, this particular review glossed over other types of limitations—including technical limitations, operational limitations, and regulatory/validation limitations—which are also well-documented challenges [132,133,134,135]. Additional reviews could be performed beyond the NASA Technical Reports Server and/or using search terms focusing on specific types of limitations to more fully understand the landscape of capabilities and limitations of AI systems.

8. Conclusions

AI and ML are powerful tools if used correctly. In non-safety-critical scenarios, it is relatively easy to apply AI models that are good at finding correlation but not necessarily causation. Assuming a causal relationship when only a correlation exists can cause AI to be overly brittle to new environments, and it is a disadvantage of AI versus previously established physics-based models. AI is difficult to implement effectively because most applications require the user to be well informed with respect to AI as well as be an expert in the application domain. Because of the advancements of NASA and others in the community, AI does not require expertise to implement. However, non-experts might gravitate toward more conventional methods instead of truly identifying which algorithm is best. Despite these potential pitfalls, AI has been successfully used in many predictive applications.

In safety-critical applications, AI becomes even more difficult to use but for different reasons. AI does not behave as a human does and can often be unpredictable. This, coupled with its brittleness to new environments, makes it difficult to verify. The successes found in safety-critical environments to this point are all examples of custom AI algorithms that target for a small task. There needs to be more work undertaken in verifying these AI machines in order for larger tasks to be entrusted to them. While the AIs are capable of many great feats, the verification techniques have not caught up to them, and so they are not used as often as they could be. Public and user trust in these systems must also be adequate, or they will not see adoption.

AI is making progress, improving results and becoming a well-known tool across many fields of study. It only appears to be growing more dominant. In non-safety-critical conditions, AI generally sees ready and wide adoption. However, in safety-critical scenarios, there are major barriers of verifying trust and trustworthiness before AI can be used more generally.