1. Introduction
Exploring and utilizing ocean resources has become a matter of strategic importance around the world against the backdrop of global resource scarcity. Underwater equipment is being increasingly investigated given the rising demands of underwater operations [
1]. The underwater vehicle–manipulator system (UVMS), which is composed of an underwater vehicle and manipulator(s), is designed to expand the capacity and enhance the functionality of underwater operations. However, the attached manipulator(s) increases the complexity of the system, which brings new challenges in terms of whole-system design, multibody kinematics and dynamics and convoluted control algorithms [
2]. Significant efforts have been devoted to developing a reliable and valuable UVMS that can operate in harsh underwater conditions.
Precise trajectory tracking control is the key technology to improve the autonomy of UVMSs. Generally, the control algorithms of UVMSs are mainly classified into two main categories. In the most straightforward sense, the control algorithms of the main vehicle and the manipulator are implemented for the corresponding degrees of freedom (DOFs) independently. In [
3], the authors combined a standard proportional–integral–differential (PID) control with a prior model to track the desired vehicle velocity, while a cascade joint controller was employed to control the velocity and angle of each joint. This approach does not consider the couplings, so the system performance is necessarily limited. In improved methods, the dynamic couplings between the vehicle and manipulator are treated as disturbances and handled by feedforward compensation [
4]. In [
5], the authors discussed the coupling property and proposed a Slotine sliding mode approach to reduce the coupling effect. Active compensation is implemented in the control law to counteract the undesirable effect by measuring or estimating the interaction force from the manipulator in [
6]. In general, the simplicity of this approach makes it attractive, and some control techniques from manipulators [
7,
8] and unmanned aerial vehicles (UAVs) [
9] can be readily referenced and transferred. However, the main vehicle will usually be designed to be much heavier than the attached manipulator(s), resulting in relatively smaller reactive forces from the manipulator(s) compared to the vehicle’s own weight, as in the Phoenix+SMART3S [
2].
Another control scheme is to model the vehicle and the attached manipulator(s) as a serial chain of rigid bodies by regarding the UVMS as a whole system. However, more challenges will arise due to the increased complexity of the whole system, such as inaccurate parameters of the hydrodynamic effects, kinematic redundancy of the system, etc.
Several classical control methodologies have been successfully applied to trajectory tracking in UVMSs, including robust PID in [
10,
11,
12,
13,
14], sliding mode control (SMC) in [
15,
16,
17,
18] and adaptive controllers in [
19,
20,
21,
22,
23], which are capable of effectively addressing external disturbances and unmodeled dynamics. Furthermore, advanced nonlinear control strategies, such as
control in [
24,
25], prescribed performance control in [
26,
27] and neural network techniques in [
28,
29,
30,
31], have also been implemented to control underwater vehicles. However, these methods demonstrate limited capabilities in handling state constraints. Consequently, model predictive control (MPC) has been extensively adopted due to its inherent constraint handling advantages in the field of underwater vehicles, as shown in [
32,
33,
34,
35,
36].
The previously mentioned works focus on tracking the trajectory based on the position, which is difficult to measure accurately in the underwater environment. Vision sensors are an important means of localization in a GPS-denied environment, which is referred to as visual servoing; this has also been widely used in underwater vehicles for dynamic positioning [
37,
38], path following [
39], docking [
40,
41] and underwater manipulation [
42,
43].
The capacity to handle interactions is one of the fundamental requirements to accomplish the manipulation task successfully [
44]. The contact force at the manipulator’s end-effector is employed to describe the state of interaction. Thus, force control strategies are often employed in scenarios where a robot intervenes in the environment, and they can be classified into ’direct force control’ and ’indirect force control’ according to whether the force is controlled directly or not [
45]. Both of them are mainly constructed based on position control schemes and have been widely used for fixed-base manipulators.
However, the velocity and position of free-floating robots are difficult to measure accurately and rapidly, making it challenging to practically implement force control algorithms for freely moving manipulators [
46], in which even higher precision and update rates are necessitated than in position control. Therefore, the current force control studies on mobile robots, including wheeled mobile robots [
47,
48,
49] and UAVMs [
50,
51], remain largely confined to laboratory environments, with limited real-world applications. The same situation applies to UVMSs, including both impedance control ([
52,
53]) and direct force control ([
54,
55,
56]).
In line with the distinction between position-based visual servoing (PBVS) and image-based visual servoing (IBVS), force control strategies can also bypass the reconstruction of the position and directly regulate contact forces using visual information. In image-based force control, the translational positions are replaced by visual features; then, visual features and forces are simultaneously tracked in a unique control law to handle the interaction with the environment. A vision-based impedance force control algorithm was introduced for a laboratory injection system in [
57]. The injection force was estimated by utilizing visual feedback via the concept of visual–force integration. Meanwhile, three types of hybrid visual–impedance control schemes that directly relate image feature errors to external forces by projecting the force component onto the image plane were proposed in [
58]. This approach enables stable physical interaction between an unmanned aerial manipulator and its task environment.
Challenges also arise from the complexity of handling multiple DOFs in the whole UVMS system. A set of tasks (SOT) is expected to be carried out simultaneously with prescribed priorities by utilizing the redundant DOFs of the UVMS. Task-based control methods for UVMSs have been proposed in [
59,
60,
61] with the consideration of priorities between tasks. In [
62], a whole-body control (WBC) framework for a dual-arm UVMS is proposed to deal with multidimensional inequality control objectives. In this work, WBC is firstly introduced for the controller of the UVMS. Although it is similar to some task-priority-based control methods, it offers a distinct perspective on the problem by utilizing the whole system’s DOFs.
Considering the advantages of IBVS, a vision-based force control task is established without the position reconstruction process. By employing image moments instead of point features, this method achieves enhanced robustness. Furthermore, a multitask framework considering smooth transitions can preserve the priority of the force tracking task while effectively mitigating fluctuations.
In our previous work, an image-based visual serving method for UVMSs was introduced [
42]. The hierarchical control architecture is composed of a kinematic model predictive IBVS controller and a dynamic velocity controller, respectively. In this work, a whole-body vision/force control framework is proposed to simultaneously track the reference vision and force trajectory, which fully exploits the whole-body DOFs of the UVMS. The contributions of this work are summarized as follows:
A whole-body multitask control framework is proposed to simultaneously accomplish the visual trajectory tracking and contact force tracking of a UVMS. This approach allows flexible task combinations in different scenarios while consistently maintaining strict priorities.
By reprojecting the image points and choosing proper image moment combinations, decoupled visual features aligned along the tangential and normal directions of the target plane can be obtained, which enhances the independence among individual tasks.
A continuous transition method for SOT switching is presented to optimize the performance in the transition process, particularly by effectively reducing fluctuations in the contact force.
This paper is organized as follows. The system formulation and visual servo model are introduced in
Section 2. Several tasks for hybrid vision/force control based on WBC are established and a hierarchical WBC framework is reviewed in
Section 3. In
Section 4, a continuous transition method for SOT switching in WBC is proposed. The simulation results are demonstrated in
Section 5. Lastly, we draw the conclusions of our paper and discuss further research problems.
5. Simulation Results
In this section, a visualized numerical simulation is conducted to evaluate the proposed controller. The system dynamics and control algorithm are calculated by MATLAB (version: R2024b), and the simulation visualization is performed by CoppeliaSim, as shown in
Figure 3. The model of the main vehicle is based on the Kambara underwater vehicle model in [
66], and a sway thruster is added to the thrust mapping matrix to render the vehicle fully actuated. The parameters of the attached manipulator can be seen in
Appendix A. The links are assumed to be symmetric cylinders to simplify the calculation of the hydrodynamic forces, and the details can be found in the SIMURV 4.1 paper [
2]. The optimization problem is solved by the open-source solver
CasADi in [
67].
The dynamic velocity controller is detailed in [
42]. To verify the robustness of the controller, we apply a time-varying external force on the end-effector of the UVMS as a disturbance and add a 10% error to the model parameters. The external force is set as
in the inertial frame. The parameters of the contact model are set as
,
and
, and Gaussian white noise of 10 db is added in the force measurement in the simulation.
Two common cases that require contact force tracking are considered in this section. One involves maintaining a contact force with a fixed point, such as button pressing. The other requires maintaining the desired contact force while moving the contact point, such as welding or flaw detection. The simulation is conducted based on these two cases.
The tasks for the hybrid vision/force control mission are listed in
Table 2. Depending on whether task
is active or not, the SOTs are defined as
and
.
5.1. Fixed Contact Case
A reference trajectory for the end-effector in the fixed contact case is considered in this part. The initial states of the UVMS are set as
and the fixed reference states are
The results of the fixed contact case based on OWBC and the proposed CWBC are shown in
Figure 4,
Figure 5,
Figure 6,
Figure 7 and
Figure 8. Overall, both of the controllers are capable of driving the end-effector to the desired point and maintaining a stable contact force.
The image moments in the tangential directions of the target plane and the end-effector quaternions are demonstrated in
Figure 4 and
Figure 5. It can be observed that both the tangential image moments and the end-effector quaternions exhibit similar performance for the two controllers.
The main difference is reflected in the normal direction in
Figure 6. After task
is introduced into the SOT at about 4 s by
, the contact force of CWBC in green converges to the reference value more quickly and stably than the red curve of OWBC. Although the force control tasks are the same in the two controllers, the better performance comes from the strictly hierarchical priorities in CWBC.
The system velocities are demonstrated in
Figure 7 and
Figure 8. It can be observed that CWBC exhibits greater fluctuations in the velocities of the main vehicle, whereas OWBC shows larger fluctuations in joint motions. This results from the fact that OWBC does not impose strict task prioritization, and its performance aligns more closely with the overall energy consumption by moving the joint in the first place.
5.2. Moving Contact Case
A circular reference trajectory for the UVMS is considered and designed as
and the other conditions are the same as in
Section 5.1. The results are shown in
Figure 9,
Figure 10,
Figure 11,
Figure 12 and
Figure 13. The contact condition between the environment and the end-effector is subjected to change in order to evaluate the controller’s adaptability to different situations during the tracking process, which is achieved by changing the equilibrium position and stiffness coefficient to
and
.
In the moving case, similar performance can be seen, as shown in
Figure 9 and
Figure 10. One can notice that there is little difference between the curves of the tangential image moments and the end-effector quaternions for the two controllers.
The normal image moment and contact force curves are shown in
Figure 11. Although both of the controllers can maintain the reference force, thanks to the strict priority of the force control task, the result of CWBC is obviously better than that of OWBC, with fewer fluctuations. At about 12 s, the normal contact forces suddenly increase when the contact condition changes and then gradually stabilize to the expected value. The adaptability of the two controllers can be certified.
Figure 12 and
Figure 13 exhibit the results regarding the system velocities. The difference between the velocities of the main vehicle and the joint velocities remains similar to that described in
Section 5.1, in which CWBC shows greater fluctuations from the main vehicle movements and OWBC shows larger fluctuations in joint motions.
5.3. Parameter Selection
The control performance of the proposed CWBC controller under different control gains in the force control task is demonstrated in this part to show the selection process of the control gains. All the initial and reference states remain consistent with those specified in
Section 5.1. The fixed contact case is repeated three times with different
: Case 1 with
, Case 2 with
and Case 3 with
.
The results of the three cases are shown in
Figure 14. It can be observed that the the performances differs notably. Specifically, Case 2 demonstrates slow tracking performance accompanied by gradual fluctuations, while Case 3 achieves fast tracking at the cost of noticeable oscillations. In general, the control force of Case 1 achieves balanced overall performance in terms of tracking speed and fluctuations.
5.4. Comparative Results
A comparison of
to
via smooth transitions with CWBC and by direct switching with HWBC is demonstrated in this part. The reference trajectory and parameters are the same as in
Section 5.2, without contact condition changes, and the results are shown in
Figure 15 and
Figure 16.
The contact forces are shown in
Figure 15. In general, both of the curves converge to the desired values stably under the two controllers, while the two methods exhibit different behaviors during the transition phase. The force result in HWBC shows larger fluctuations, whereas CWBC approaches the desired value more smoothly. Significant fluctuations are observed in the velocity along the force control direction in
Figure 16, which is consistent with the variation in the contact force. This demonstrates the effectiveness of the smooth transition method in CWBC to optimize the switching performance when the force control task is introduced in the SOT.
The root mean square error (RMSE), settling time (ST) and overshoot (OS) of the two controllers are demonstrated in
Table 3. The RMSE and the ST are relatively close, indicating that the convergence rates of the two controllers are similar. A notable distinction between the two controllers is observed in the OS. The OS of CWBC is 1.8 (N), while the OS of HWBC is 8.1 (N). This is not only reflected in the improved tracking performance but is particularly critical in intervention tasks, as excessive overshoot of the contact force could potentially lead to system damage or instability.