Data-Driven Selection of Decontamination Robot Locomotion Based on Terrain Compatibility Scoring Models

Abstract

Decontamination robots are becoming more common in environments where reducing human exposure to hazardous substances is essential, including healthcare settings, laboratories, and industrial cleanrooms. Designing terrain-capable decontamination robots quickly is challenging due to varying operational surfaces and mobility limitations. To tackle this issue, a structured recommendation framework is proposed to automate selecting optimal locomotion types and track configurations, significantly cutting down design time. The proposed system features a two-stage evaluation process: first, it creates an annotated compatibility score matrix by validating locomotion types against a robust dataset based on factors like friction coefficient, roughness, payload capacity, and slope gradient; second, it employs a weighted scoring model to rank wheel/track types based on their appropriateness for the identified environmental conditions. User needs are processed dynamically using a large language model, enabling flexible and scalable management of various deployment scenarios. A prototype decontamination robot was developed following the proposed algorithm’s guidance. This framework speeds up the configuration process and establishes a foundation for more intelligent, terrain-aware robot design workflows that can be applied to industrial, healthcare, and service robotics sectors.

1. Introduction

Decontamination removes oil or chemical substances from objects, surfaces, or environments, reducing the risk of contamination and accidents [1]. Cleaning oil or grease-stained floors requires specific cleaning products and methods such as scrubbing, brushing, and pressure washing, as shown in Figure 1 [2]. Manual or traditional decontamination processes, including segregation, cleaning, disinfection, and sterilization, are often labor-intensive, time-consuming, and inconsistent [3,4]. Prolonged exposure to harsh chemicals in cleaning solutions may cause skin irritation, respiratory issues, or allergic reactions [5]. Inhaling fumes from potent cleaning agents, especially in poorly ventilated areas, can lead to respiratory distress, dizziness, or headaches [6]. Additionally, the physical act of scrubbing or pressure washing, especially when using high-pressure machines, can lead to musculoskeletal injuries such as strains or sprains due to repetitive motion or incorrect posture [6].

Robots assist in decontamination by automating cleaning tasks, reducing human exposure to hazardous chemicals and pathogens [7]. Decontamination includes pressure washing, brushing, UV-C light, chemical spraying, and steam. Using robots minimizes risks to humans from contact with toxic substances, improving safety, and ensuring effective, consistent cleaning [8,9]. Significant features of a decontamination robot include locomotion, autonomous navigation, decontamination methods, obstacle detection, and real-time monitoring. It has sensors for precise mapping, contamination detection, and data logging. Safety features ensure protection from harmful substances, while user interfaces allow for easy operation [9]. Locomotion is crucial for a decontamination robot as it enables efficient movement across different surfaces, ensuring complete coverage [10]. Whether utilizing wheeled, tracked, or omnidirectional systems, the robot’s mobility enables it to navigate complex environments, access hard-to-reach areas, and execute thorough cleaning, thereby enhancing the effectiveness and versatility of the decontamination process, as presented in Figure 2. These decontamination robots are increasingly employed in healthcare facilities, cleanrooms, public spaces, and warehouses to reduce human exposure to contaminants and improve cleaning efficiency [1,2,5,8].

Locomotion in decontamination robots is crucial for navigating environments efficiently. It ensures stability on uneven surfaces and improves operational flexibility. Wheeled robots are common, offering quick, flexible movement ideal for cleaning, inspection, and navigation in organized spaces. Generally, the active locomotion wheels of the robots can be used to differentiate the type of locomotion, such as differential drive, skid-steering, Ackermann steering, or omnidirectional configurations, as illustrated in Figure 3. Omnidirectional locomotion robots represent a subset of multi-wheeled locomotion, employing wheels that permit movement in any direction without orientation changes. This capability offers them high maneuverability, making them perfect for tasks that require precise navigation in tight or congested spaces, including wall climbing, sweeping, or cleaning in environments with minimal obstacles. An additional aspect of wheeled locomotion is the rocker-bogie suspension system utilized in robots, which must uphold stability on challenging or uneven terrains [11]. It consists of a set of connected wheels that enable the robot to adjust its body orientation, enhancing traction and stability over obstacles.

Tracked locomotion robots are the second most used robotic system category, offering better traction and stability on rough terrains compared to wheeled robots. As shown in Figure 4, they are classified into six types based on track design: single, dual, triangular, articulated, hybrid, flexible belt, and rigid metal tracks. Each type provides unique benefits in stability, maneuverability, and terrain adaptability, making them ideal for rugged environments. Commonly used in industry, search and rescue, and challenging terrains, they outperform wheeled robots on difficult surfaces.

Designing decontamination robots involves choosing locomotion, navigation, and decontamination methods to operate efficiently in various environments. Challenges include customizing robots for specific tasks like cleaning surfaces or using different disinfection methods. To overcome these challenges, several researchers [12,13,14,15,16] explored the use of advanced artificial intelligence (AI) models, large language models (LLMs) like chat generative pre-trained transformer (ChatGPT 4o) and natural language processing (NLP), and other machine learning techniques, which leverage vast datasets of robot specifications. By processing user inputs like surface type, disinfection method, or robot size, these models analyze and select the best robot design for specific decontamination needs. Researchers are refining this technology to improve accuracy and adaptability, streamlining robot design and enabling easier adaptation for complex tasks in various settings.

The remainder of this paper is structured as follows. Section 2 presents a comprehensive review of related studies, including robotic decontamination methods as well as various locomotion strategies. Section 3 introduces the proposed data-driven locomotion recommendation algorithm. Section 4 details the design of the decontamination robot. Section 5 demonstrates the prototype development of the designed decontamination robot by highlighting limitations and future scope in the proposed design. Section 6 concludes the study by summarizing the main contributions and findings.

2. Related Studies

This section examines studies on decontamination robots, emphasizing the challenges of selecting suitable techniques. It highlights different locomotion methods used in various robots for specific applications. It addresses advancements in AI-driven models to analyze user inputs and optimize robot selection for tasks and environments.

2.1. Decontamination Robot

Petereit, Janko et al. developed robotic systems for decontamination in hazardous environments, focusing on nuclear decommissioning, plant component decontamination, and waste handling [1,2]. It involves research on mobile robots, autonomous machines, 3D mapping, and teleoperation, with practical applications aimed at improving safety and reducing health risks. Liu, Shengyong et al. reviewed various decontamination methods, highlighting laser-based cleaning, chemical gels, and their potential for improving efficiency. It emphasizes combining methods for optimal decontamination results and calls for further development of versatile, effective solutions [8]. Al-Dubooni et al. developed a hybrid continuum-eversion robot for remote decontamination in the nuclear industry, combining flexibility and precision. It successfully delivers liquids and aerosols in hard-to-access areas, achieving over 95% success in spraying tests [9]. Miura Riku et al. developed a remotely operated robot for decommissioning that features mecanum wheels, crawlers, and an expandable ladder system to access inaccessible areas. Despite challenges and malfunctions, it demonstrated adaptability for future nuclear plant decontamination [10].

2.2. Locomotion

In robot locomotion, wheeled locomotion is commonly used over other methods due to its simplicity, speed, and efficiency on flat surfaces, offering lower energy consumption, reduced complexity, and cost-effectiveness. In wheeled locomotion, types like ball [17], two-wheel [18], three-wheel [19], and multi-wheel [20], each with drive mechanisms like differential drive, Ackermann steering [21], and omnidirectional, serve different applications as illustrated in Figure 5. Ronald Ping Man Chan et al. [18] investigated the constraints of maintaining stability for a two-wheel robot. The two-wheeled robot was most suitable for flat terrains. Javier Moreno et al. [19] developed a three-wheeled holonomic motion system with autonomous navigation capabilities for navigating complex, even indoor terrain. Ackermann steering is a standard method of locomotion for wheeled vehicles. It enables the front wheels to steer at different angles during turns, making it suitable for indoor and outdoor paved surfaces [21]. Differential drive is mainly used for robotic applications, where the two/four wheels must be controlled independently, allowing the vehicle to move in different directions by varying the speed of each wheel. Charan Vikram et al. [20] developed an omnidirectional floor-cleaning wheeled robot equipped with four omnidirectional wheels for narrow and crowded spaces. Jianfeng Liao et al. [22] designed a four-wheel independently driven skid-steered mobile robot designed for industrial automation and outdoor exploration. It addresses the chattering phenomenon in skid-steered robots by proposing a coordinated adaptive robust control, ensuring smooth operation and improved control performance under varying ground conditions. The synchro-drive principle relies heavily on gear mechanisms, where one motor controls the rotation around the horizontal axis, generating traction, and another controls the rotation around the vertical axis for direction [23]. Articulated pivot steering connects the front and rear of heavy machinery via a central pivot joint, allowing independent movement that reduces the turning radius and enhances maneuverability in confined areas [24].

Tracked locomotion offers improved traction, stability, and navigation on rough terrain, making it well-suited for heavy-duty applications. As shown in Figure 6, tracked robots include types like continuous track, flexible belt, triangular track, flipper track, rigid metal track, articulated, dual, and hybrid, each tailored for specific terrains. Continuous track robots, also known as caterpillar track robots, use a single band of treads and excel in rough or soft terrain, offering superior traction and stability compared to wheels [25]. A dual-track robot uses two tracks for stability and traction on rough terrain. It is ideal for rugged landscapes, construction sites, and military operations, where mobility and load-bearing are essential [26]. An articulated track robot features a flexible design with two segments connected by a pivot, allowing for improved maneuverability in tight spaces. It is ideal for narrow or cluttered environments like disaster response, search and rescue, and agricultural applications [27]. A triangular track robot utilizes a three-pointed system for traction and stability on rough terrain, making it ideal for steep or uneven surfaces, such as mountains, construction sites, and exploration areas. A staircase-climbing robot navigates stairs and vertical obstacles, making it useful in indoor environments [28]. Wenzhi Guo et al. [29] developed TALBOT, a tracked-leg transformable robot designed for all-terrain adaptation [30]. Flexible belt track offers mobility on soft or uneven terrain, which is suitable for agricultural fields, forests, and muddy or sandy areas needing smooth movement [31]. Flipper tracks have flipping mechanisms for overcoming obstacles in highly uneven terrain, rocks, or boulders, and are often used in search and rescue or space exploration [32]. Rigid Metal Track ensures durability and traction on rocky or icy surfaces, utilized in military, mining, and heavy-duty industrial applications demanding extreme durability [33].

Hybrid locomotion robots mimic biological movement and are designed for environments unsuitable for wheels or tracks, like uneven terrain, stairs, or vertical surfaces. With multiple joints, they offer flexibility and adaptability to complex environments (see Figure 7). These robots combine various locomotion types, wheels, legs, and tracks, enabling them to switch movement styles based on the terrain. Hybrid systems are ideal for applications needing versatility across both flat and rugged surfaces. A snake robot is a flexible, articulated robot that mimics snake movement with multiple segments for undulating motion. It is ideal for navigating tight spaces like pipes, ducts, and disaster zones where traditional robots cannot operate [36]. Soft robots are flexible, deformable robots made from soft materials, enabling adaptation to complex environments. Using soft actuators (like pneumatic or hydraulic systems) for movement, they excel in fragile settings and unstructured terrains like underwater or narrow spaces. Their adaptability allows them to handle delicate objects and navigate confined areas [37]. Takeru Yanagida et al. [38] developed Scorpio, a bio-inspired, reconfigurable robot with rolling, crawling, and wall-climbing abilities for urban reconnaissance. A bipedal robot mimics human walking with two legs. It is built for balance and mobility on various surfaces. Biped robots excel at human-like tasks, including service roles and exploration [39]. Xuejian Qiu et al. [40] proposed an autonomous wheeled-legged robot system that can traverse uneven terrain efficiently. Qimeng Li et al. [41] researched on Quadruped robots, a modular, robust quadruped robot designed for harsh environments, featuring dynamic motion and terrain mapping.

A comprehensive robot locomotion system, well-suited for various terrain types, is summarized in

Table 1. Complete robot locomotion system suited for different terrain types.

Terrain Type	Recommended Locomotion Type	Subtype/Example	Justification
Flat Indoor Surfaces	Wheeled Locomotion	Differential Drive, Ackermann Steering	Efficient, fast, low-energy use on smooth, predictable surfaces
Confined/ Cluttered Areas	Omni-Directional Wheels	Mecanum Wheels, Swedish Wheels	Enhanced maneuverability, zero turning radius
Sandy/ Uneven Outdoor	Tracked Locomotion	Flexible Belt, Flipper Track	Greater ground contact, better traction, and stability
Rocky/ Rugged Terrain	Hybrid (Wheel-Leg, Tracked-Leg)	Rocker-Bogie, Transformable Track-to-Leg Mechanism	Combines mobility and terrain adaptability
Pipeline/ Ventilation	Magnetic Wheeled, Snake-like Robots	Multi-link Magnetic Wheels	Adaptability to curved/narrow environments, wall climbing capability
Stairs/ Elevations	Articulated or Triangular Tracked	Triangular Track with Stair-Climbing Kinematics	Designed to overcome vertical height changes and obstacles
Urban Debris/ Disaster	Legged or Wheel-Leg Hybrid	Biped, Hexapod, Scorpio (rolling + crawling + climbing)	Ability to navigate unpredictable, obstructed, or destroyed environments
High-Slip/ Slippery	Skid-Steering with Sensor Feedback	4-Wheel Skid-Steer Robot	Tolerant to wheel slip, simple mechanism, suited with proper slip control logic
Cleanroom/ Laboratories	Wheeled or Synchro-Drive	Synchro-Drive with Sterile Design	Precise control, low noise, clean-surface compatible

. Overall, these robots inspired the dataset preparation, where wheeled, tracked, legged, and hybrid locomotion systems each serve distinct purposes and applications. Wheeled robots offer speed and efficiency on flat surfaces, tracked robots excel in rough terrains, legged robots provide flexibility in complex environments, and hybrid systems combine multiple locomotion types for versatile performance across various tasks.

2.3. Large Language Models

AI-based textual models like LLMs are designed to process and generate human-like text based on text-based input prompts. These models are increasingly being integrated into various fields, including design and robotics. For instance, Zhe Zhang et al. introduced Mani-GPT, an interactive robotics model that interprets human instructions and generates appropriate plans for robotic manipulation, achieving 84.6% accuracy in intent recognition and task execution for pick-and-place robots [12]. Similarly, Qihao Zhu et al. [13] explored the use of NLP and GPT technologies to automate the early design stages of computer-aided design (CAD), where textual prompts are converted into new design concepts through a three-step process: domain understanding, problem-driven synthesis, and evaluation-driven synthesis. Meanwhile, Christos Konstantinou et al. applied Gen-AI in human-robot collaboration within industrial settings, utilizing behavior trees and LLM technologies to enable real-time adaptability in assembly tasks, integrating blockchain instructions for enhanced reliability [14]. Likewise, Adam Fitriawijaya et al. used ChatGPT and multimodal generative AI (Gen-AI) for architectural design, incorporating the Ethereum blockchain to enhance the authenticity, traceability, and ownership of the data used in the design process [16]. Lastly, Sai H Vemprala et al. developed PromptCraft, an open-source tool designed to generate prompts for various robotic applications, integrating ChatGPT to improve simulations and task performance in free-form dialogues, geometric navigation, and aerial tasks [15].

While numerous studies concentrate on nuclear decontamination methods, research into robots for decontamination tasks is sparse. Moreover, a substantial gap exists in studies that focus on the design of decontamination robots, particularly concerning locomotion systems tailored to specific environments and tasks. Current research on locomotion for decontamination robots is insufficient, with few studies addressing the challenges of navigating complex environments while maintaining efficiency. Furthermore, no studies investigate the integration of LLMs to personalize robot locomotion based on task requirements and environmental conditions.

The research objectives considered in this study are categorized into three segments:

• Identifying the optimal locomotion system for decontamination robots based on task requirements and environmental conditions to ensure efficiency in various settings.
• Integrating with LLMs to generate a best-fit design for the decontamination robot based on user inputs and environmental factors.
• Developing the decontamination robot based on the best-fit design produced by the LLM integration, implementing the optimal locomotion system, decontamination methods, and safety features for autonomous operation in complex environments.

This integration has the potential to enhance the adaptability of robotic systems, thereby facilitating the dynamic optimization of locomotion systems tailored for various decontamination tasks. This, in turn, significantly improves their effectiveness and operational flexibility across diverse settings.

3. Proposed Data-Driven Locomotion Recommendation Algorithm

The proposed system is developed by collecting data from 20 robots whose capabilities are closely related to decontamination tasks, including fumigation robots, agricultural robots, and floor-cleaning robots. These robots share comparable operational environments and design requirements, including payload capacity, terrain type, slope navigation, and decontamination mechanisms. A complete summary of the dataset used is included in the Supplementary File. An LLM then trains on this dataset to learn the relationships between tasks, environmental conditions, and optimal locomotion types. The LLM used in this study was not fine-tuned in the conventional supervised-learning sense. Instead, it was employed in a prompt-based inference mode using the robot dataset provided in the Supplementary File. This allowed rapid customization and specification matching based on user-input task requirements. While the LLM selection was based on its broad generalization capability and context understanding, this work focuses on demonstrating its utility in design inference rather than benchmarking its training pipeline. When users provide task specifications, such as environment type, decontamination method, and robot size, the LLM processes these inputs to find the best-fit locomotion. Automating the selection process with the LLM significantly reduces design time, removing the need for manual iteration. The robot is developed using LLM-generated specifications and validated through real-world testing to ensure effectiveness in diverse environments.

In order to systematically ascertain the most appropriate locomotion and track configuration for a robot, a methodical evaluation process is proposed, contingent upon environmental conditions. As depicted in the flowchart presented in Figure 8, the procedure commences by extracting essential operational parameters from the user’s prompt utilizing a language model. These parameters are subsequently employed in a two-stage evaluation: initially, validating general locomotion types (wheeled or tracked) across various terrains; and subsequently, assessing specific track designs based on their mobility characteristics and environmental adaptability. Compatibility assessments and score assignments are executed iteratively, with the final recommendations being visualized through an annotated score matrix and bar charts to facilitate informed decision-making.

The proposed Algorithm 1 starts by taking a natural language prompt from the user (user_prompt) that details the terrain, expected payload, friction, and slope. It uses the Deepseek-R1 model to parse this prompt into an LLM_Response, extracting a structured Parsed_JSON object with relevant fields. The LLM model breaks down the user’s instructions into terrain_data, which includes Terrain Type, Payload capacity, and Slope. These details are categorized into terrains with properties like μ, roughness, slope, and weight, alongside locomotion_specs that define the operational limits for each locomotion type. The first sub-routine of Algorithm 1 is dedicated to validating various locomotion types (such as wheeled or tracked bases) by comparing their operational limits with a set of terrain parameters specified in the dataset. Compatibility for each terrain-locomotion pairing is assessed (score[locomotion]), and reasons for any failures or successes are recorded in a matrix that includes μ, roughness, and slope. Scores are determined based on successful matches, with higher importance terrains receiving greater weights.

In the second sub-routine, during the Track-Type Recommendation phase, the track_dataset is loaded with various track designs and specifications. A new score[track_type] dictionary is created to assess each track’s performance. Points are awarded in the scoring system based on the alignment of the track’s Payload Capacity (kg), Min Friction (μ), and Climb Angle (deg) with the user’s requirements for payload, friction, and slope, with partial penalties for near-misses. Furthermore, matching the terrain to Suitable Environments yields bonus points. Tracks are ranked in descending order of their scores, and a bar chart visualization is produced to illustrate their suitability. The final outputs feature a seaborn-generated score matrix for locomotion compatibility, along with a matplotlib bar chart showcasing scored track recommendations, which helps the user make an informed selection.

Algorithm 1: To Propose a Detailed Locomotion Type based on the User’s Requirement

Input: 1. Input: user_prompt←Prompt with terrain, payload, slope, and friction. E.g., “Suggest me the right locomotion type for a decontamination robot that can navigate in <terrain> with a <payload in kg> payload and a maximum slope of <degrees of slope>.” 2. LLM_Response←Call LLM prompt to extract JSON from user_prompt 3. Parsed_JSON←extract JSON fields: terrain, payload, slope, friction

Purpose:   Part A: Validate mobile base (wheel/track locomotion) against all terrains   Part B: Score specific track types based on user’s environment

Compare:   A: locomotion limits (μ, payload, roughness, slope) vs. all terrain properties   B: user requirements vs. track_type capability specs

Annotate:   A: terrain-locomotion matrix with reasons: “μ”, “P”, “R”, “S”, or “✓”   B: assign weighted scores for each track type (maximum 10)

Pseudocode: Sub-Routine A: Mobile Locomotion Evaluation 4. Load: terrain_data←list of terrain dicts with μ, roughness, slope, weight, type 5. Load: locomotion_specs←dictionary with payload, μ, slope, roughness limits 6. Initialize: score[locomotion]←0 for each locomotion type 7. Scaled Penalty: Let Pu be the user-specified payload and Pt be the track type’s maximum payload capacity. Similarly, for terrain, terrain roughness, and slope. 8. The compatibility score is updated as $\mathrm{Score}+=\left\{\begin{array}{c}2 if P_u\le P_t\\ 2\times \left(1-{\displaystyle \frac{P_u-P_t}{P_u}}\right), if P_u>P_t\end{array}\right.$ 9. For each locomotion_type in locomotion_specs: 10.   For each terrain in terrain_data: 11.     Check compatibility: 12.       If terrain_μ < loco_μ_min → append “μ” 13.       If payload > loco_payload_limit → append “P” 14.       If terrain_roughness > max_roughness → append “R” 15.       If slope > max_slope_deg → append “S” 16.     If no fails: 17.       Append “✓”, add terrain[“Weight”] to score[locomotion] 18.     Annotate terrain matrix with reasons Sub-Routine B: Wheel/Track-Type Recommendation 19. Load: track_dataset←DataFrame with track types and attributes 20. Initialize: score[track_type]←0 for each track 21. For each track_type in track_dataset: 22.   score += 2 if payload ≤ track[“Payload Capacity (kg)”] else scaled penalty 23.   score += 2 if friction ≥ track[“Min Friction (μ)”] else scaled penalty 24.   score += 2 if slope ≤ track[“Climb Angle (deg)”] else scaled penalty 25.   score += 4 if terrain ∈ track[“Suitable Environments”], 2 if partial match 26. Rank track_types by descending score 27. Generate bar chart of scored track recommendations

Output: 28. Figure–Score matrix of locomotion compatibility vs. terrain (annotated) 29. Figure–Score list and printed reasons for locomotion suitability 30. Figure–Bar chart of track-type scores (threshold-colored)

4. Design of Locomotion System of the Decontamination Robot

This section builds on the data-driven locomotion recommendation algorithm discussed in the previous section, emphasizing its practical application in designing the Decontamination Robot. It details the specific environments where the robot will operate and identifies the test surfaces chosen for assessment. Furthermore, it explains how the locomotion recommendation system processes natural language prompts that describe real-world conditions to determine the best mobility configurations.

4.1. Identifying the Locomotion System

A varied range of test surfaces was analyzed to determine the best locomotion strategy for a decontamination robot, as illustrated in Figure 9. This selection includes polished concrete indoors, epoxy-coated hallways, wooden-concrete transition areas, and outdoor inclined terrains, typically found in commercial spaces, building entrances, and landscaped pathways. A natural language description of these surfaces, alongside details on payload, coefficient of friction, and moderate slope, was supplied to the LLM-based algorithm. The model interpreted the prompt and assessed the compatibility of the terrain with the robot numerically, utilizing a multi-criteria validation framework.

4.2. Response from the Algorithm

The proposed algorithm for locomotion recommendation generates the output depicted in Figure 10. This output is derived after analyzing the user’s prompts in conjunction with the dataset. The score matrix visualization presented in Figure 10 explains the compatibility of seven locomotion types across eighteen terrain types. Each cell is annotated to indicate whether the locomotion was successful (✓) or failed due to one or more constraints, including insufficient friction (μ), payload overload (P), excessive surface roughness (R), or slope limitations (S). The scoring matrix was generated based on heuristic matching and expert knowledge from the compiled robot dataset. While no traditional validation dataset was used, the complete dataset is provided as a Supplementary File to facilitate reproducibility and further validation by future studies. In the desired environment, tracked robots exhibited the most suitability, demonstrating compatibility across nearly all surface types, which encompass rugged and sloped areas. A leg wheel combination mechanism also yielded favorable results; however, it encountered failures on stairs and extremely rough surfaces due to challenges associated with roughness and slope constraints. Wheeled robots performed well on polished and indoor surfaces but struggled on uneven terrain due to slope and roughness. Conversely, dynamic legged, snake, and hopper robots were disqualified on almost all surfaces, primarily due to payload limits and slope instability, culminating in a compatibility score of zero.

Each locomotion type was assigned a cumulative score based on weighted terrain compatibility. Tracked robots scored highest (15.1) and wheeled robots (10.7), confirming tracked mobility as the most robust base for decontamination across complex indoor-outdoor transitions.

Figure 11 illustrates a threshold-colored bar chart that compares various track types to enhance the design decision-making process. The Dual Track configuration achieved better performance than all other configurations, with a score of 8.0, while the flexible, hybrid, and articulated tracks also demonstrated significant adaptability. These findings corroborate that a tracked base, particularly one with dual tracks, provides the most versatile and stable foundation for deploying decontamination systems in environments characterized by mixed flooring, liquid spills, and variations in terrain.

In this section, the design process for the decontamination robot is presented, guided by a data-driven locomotion recommendation algorithm. Various test surfaces are analyzed to determine the most suitable mobility configuration. The LLM-based system evaluates locomotion compatibility and generates an annotated score matrix and score-based recommendations by employing natural language prompts that describe terrain conditions. Tracked locomotion, particularly the dual track configuration, emerges as the most suitable choice due to its high compatibility across diverse terrains. Additionally, the section explores terrain-specific constraints such as slope, roughness, and friction, thereby providing a comprehensive framework for selecting the robot’s mobility architecture. This foundational design assessment directly informs the mechanical and structural decisions made during the subsequent development phase.

5. Developed Decontamination Robot Prototype

The dual-track locomotion method has been adopted based on the proposed algorithm’s output. This section outlines the development of decontamination robots, from conceptual design to physical prototyping. A detailed 3D CAD model (shown in Figure 12 and Figure 13) was created to visualize and enhance the mechanical structure using the mobility recommendations and environmental constraints identified in earlier sections. This CAD design allowed for assessing component placement, weight distribution, and spatial limitations for onboard systems, including the locomotion system, water tanks, and control electronics. After several design iterations, a functional prototype was produced and assembled, demonstrating the proposed architecture’s viability for use on different decontamination surfaces.

5.1. CAD Modeling

The developed 3D CAD model provides a detailed visualization of the decontamination robot’s subsystem layout, highlighting the integration of key functional components. At the top of the chassis, a 3D LiDAR unit is mounted to enable real-time environment mapping, obstacle detection, and autonomous navigation. Adjacent to LiDAR, the robot houses an industrial PC responsible for processing sensor data and executing high-level control algorithms. The design incorporates dual water tanks, positioned centrally on the chassis to optimize weight distribution and maintain stability during movement. The robot’s locomotion system is driven by a dual-track mechanism, clearly illustrated on both sides of the model, ensuring improved traction and the ability to traverse uneven or inclined surfaces effectively. Five rotating brushes are installed at the robot’s front section, driven by individual motors, and strategically aligned to maximize ground coverage during cleaning operations.

5.2. Developed Decontamination Robot

The developed decontamination robot (shown in Figure 14) is a heavy-duty tracked platform engineered explicitly for rigorous cleaning tasks across various terrains. It features 230 mm wide rubber tracks that enhance grip and stability on surfaces inclined up to 30°. With chassis dimensions measuring 2000 × 1300 × 710 mm, the robot weighs 450 kg without the batteries and can carry a total load of 1200 kg, encompassing the entire payload. The robot is powered by two AC three-phase motors, each generating 2 kW, enabling a maximum speed of 2 km/h. Power is supplied by two 48 V, 200 Ah LiFePO₄ batteries, one designated for movement and the other for the cleaning apparatus. For decontamination, the robot is outfitted with five rotating nylon-bristle brushes selected for their durability and resistance to abrasion in harsh conditions. Each brush is driven by a BLMR6400SK-GFV-B brushless motor (Sourced from Oriental Motor Singapore, Singapore), which provides reliable torque and is characterized by a long service life. Furthermore, it comprises a dual-tank system: one tank for water to assist in dislodging debris and another for waste storage. The mechanical and electrical design of the robot adheres to the guidelines specified in the track selection algorithm, ensuring adaptability to diverse terrains, prolonged durability, and effective cleaning capabilities suitable for both indoor and semi-outdoor environments. A specification summary of the robot is presented in

Table 2. Developed Robot Specification Table.

Parameters	Specifications	Parameters	Specifications
Track Material	Rubber	Driving Motor (Each)	2 kW
Track width	230 mm	Maximum Speed	2 km/h
Dimensions	2000 × 1300 × 710 mm	Maximum Slope	≤30°
Self-Weight	450 kg (Excluding battery)	Battery	LiFePO4 48 V, 200 Ah
Loading Capacity	1200 kg (Including robot weight)	Brush Type	Nylon bristles

.

5.3. Limitations and Future Scope

Although the proposed algorithm has demonstrated effectiveness across the majority of use cases, there exist minor errors that prevent the algorithm from functioning optimally. Some of the limitations are as follows:

• LLM Output Deviation: The LLM is currently used in a prompt-based inference mode without fine-tuning. As such, it may return outputs with missing fields, non-JSON formatting, or unexpected data types. A possible solution is to address this by implementing robust validation schemas and retry mechanisms, as well as refining prompt engineering techniques. Fine-tuning the LLM with more structured examples may also reduce inconsistency.
• Approximate Terrain Matching: The use of partial string matching (e.g., substring or fuzzy match) between user input and pre-defined terrain types can yield suboptimal results, especially for ambiguous or uncommon terrain descriptions. A possible solution is to use a semantic similarity model, which can improve contextual matching. Additionally, expanding the terrain dataset with synonyms and example scenarios will enhance accuracy.
• Fixed Thresholds: Hard-coded thresholds for payload, slope, and friction make the system rigid and less responsive to user-specific priorities. A possible solution is to add adaptive thresholding using rule-based scaling or reinforcement learning, based on user feedback and mission history, which can enable more dynamic, context-aware decision-making.
• No Multi-Terrain Aggregation: The current system architecture evaluates one terrain type at a time, which limits its utility in multi-environment deployments. A possible solution is to incorporate a multi-label terrain evaluation system, which can assess the feasibility across multiple terrains simultaneously and recommend hybrid locomotion solutions accordingly.
• Lack of Adaptability to Unstructured Terrain: The model is primarily trained on structured or semi-structured environments found in existing robot datasets. A possible solution is to incorporate data from field robotics and search-and-rescue robots operating in dynamic terrains, combined with few-shot learning, which can improve generalization to novel terrains.

The future scope of the proposed work includes the following:

• Cross-Platform Compatibility: Design an algorithm that is capable of functioning across various robotic platforms with minimal modifications.
• Explainable Recommendation Reports: Automatically generate comprehensive explanation reports that summarize the rationale behind each locomotion or track recommendation, thereby enhancing transparency for engineers and operators.
• Environment Clustering and Unseen Terrain Generalization: Utilize unsupervised learning techniques to cluster similar terrains and extrapolate recommendations, even when encountering terrains not included in the original terrain data list.
• Dynamic Learning with Feedback: Following deployment, gather operational data on the success or failure of recommendations and subsequently retrain the scoring system or recalibrate thresholds automatically.
• Reconfigurable Locomotion: Designing modular robots with transformable locomotion enables real-time terrain adaptation, enhancing versatility in complex decontamination environments.
• Hybrid Intelligence: Integrating ontology and LLMs enables adaptive, structured robot design for diverse decontamination tasks and dynamic environments.

6. Conclusions

The proposed recommendation system provides a systematic methodology for selecting the most suitable locomotion types and track mechanisms, tailored to environmental and operational requirements, including terrain, payload capacity, surface friction, and slope. The system generates an annotated score matrix that describes feasible areas and potential operational risks by analyzing terrain data in conjunction with locomotion specifications. Furthermore, a wheel/track dataset supports a secondary, weighted scoring framework that evaluates various track designs based on their adaptability and performance. The system dynamically extracts user requirements through LLM-assisted prompt parsing, offering flexibility across a spectrum of mission profiles without reliance on hardcoded input assumptions. A prototype decontamination robot was developed to assess system effectiveness using the algorithm’s proposed locomotion and tracking configurations. The mobility performance closely aligned with the algorithm’s predictions, reinforcing the recommendation model’s efficacy. This integrated decision-making framework automates critical aspects of mobile robot design, ensuring scalability for various environments. Future enhancements may encompass the integration of user-prioritized constraint weighting, re-evaluating decisions utilizing real-time sensory feedback, and implementing simulation-validated recommendation cycles to refine the decision-making process. Collectively, the framework presents a robust, modular, and scalable approach to the design of terrain-aware robotic mobility systems, facilitating broader application in intricate field robotics scenarios.