With the long-term on-orbit operation of space stations, extravehicular status monitoring and surface inspection tasks have imposed higher demands on the motion planning capabilities of manipulators. Space redundant manipulators, when performing visual monitoring tasks, must simultaneously satisfy multiple complex constraints including kinematics, dynamics, obstacle avoidance, and end-effector line-of-sight, making this a typical nonlinear non-convex optimal control problem. Although sampling-based methods can rapidly generate collision-free paths, they struggle to guarantee dynamic feasibility in high-dimensional systems and often yield poor-quality paths. Optimization-based methods, while capable of directly generating high-quality trajectories, are highly sensitive to initial solutions and lack effective warm-start mechanisms. Therefore, how to merge the rapid search capability of sampling-based methods with the precise solving advantage of optimization-based methods to plan high-quality executable trajectories under complex constraints for space manipulators has become a key challenge in advancing the intelligent development of space robotics.
In a recent study published in Space: Science & Technology, the team led by Zhu Zhanxia from the Unmanned Systems Research Institute, Northwestern Polytechnical University, proposed a hierarchical motion planning method that integrates goal-biased Rapidly-exploring Random Tree (RRT) with sequential convex programming. The study first designs a goal-biased RRT algorithm at the kinematic level, employing a biased sampling strategy and greedy pruning to rapidly generate collision-free paths, and then constructs a smooth initial reference trajectory via cubic spline interpolation to provide a high-quality warm start for sequential convex programming. Subsequently, at the dynamic level, the obstacle avoidance constraints are transformed into linear inequalities on angular velocities using the velocity damping method; the line-of-sight constraints and dynamic equations are linearized via quaternion differentiation and first-order Taylor expansion, and the accuracy of convex approximation is ensured through successive linearization and trust-region constraints, thereby converting the original nonlinear non-convex problem into an iteratively solved convex optimization problem. Simulation results for a 7-degree-of-freedom space redundant manipulator in both static and dynamic target monitoring scenarios demonstrate that the proposed method successfully plans smooth trajectories satisfying obstacle avoidance, end-effector line-of-sight constraints, and joint torque limits, converging in 17 and 18 iterations for the static and dynamic scenarios, respectively. This hierarchical planning method effectively overcomes the poor scalability of sampling-based methods under complex constraints, while providing a high-quality initial trajectory for sequential convex programming, offering a feasible technical solution for autonomous motion planning of space manipulators under end-effector task constraints such as extravehicular visual monitoring.
First, this study focuses on the complex motion planning problem faced by space redundant manipulators when performing extravehicular visual monitoring tasks, and proposes a hierarchical planning framework that integrates goal-biased RRT with sequential convex programming. Space redundant manipulators operating in harsh environments characterized by microgravity and large temperature differentials must simultaneously satisfy multiple constraints, including kinematics, dynamics, obstacle avoidance, and end-effector line-of-sight, making this a typical nonlinear non-convex optimal control problem. While sampling-based methods can rapidly search for collision-free paths, they struggle to guarantee dynamic feasibility; optimization-based methods, although capable of generating high-quality trajectories, are highly sensitive to the initial solution. To address this, the study designs a hierarchical planning architecture as shown in Fig. 1: at the kinematic level, a goal-biased RRT algorithm is employed to generate collision-free paths, which are then smoothed via cubic spline interpolation to obtain a smooth initial reference trajectory; at the dynamic level, all constraints are convexified to construct sequential convex programming subproblems, and the optimal trajectory is obtained through iterative solution.
Second, the study elaborates on the initial trajectory generation method and the constraint convexification strategy in detail. At the kinematic level, the goal-biased RRT employs a biased sampling strategy to accelerate search convergence, utilizes axis-aligned bounding boxes in conjunction with the Separating Axis Theorem for collision detection, and eliminates redundant waypoints through a greedy pruning strategy. Fig. 2 presents the joint angle, angular velocity, angular acceleration, and driving torque profiles of the generated initial trajectory, demonstrating that the trajectory is smooth and continuous while satisfying obstacle avoidance constraints. Fig. 3 shows a comparison of computation time statistics between GB-RRT and standard RRT over 100 path planning experiments, indicating that the proposed GB-RRT achieves higher computational efficiency and can more rapidly provide a warm-start reference for sequential convex programming. At the dynamic level, the study adopts the velocity damping method to transform obstacle avoidance constraints into linear inequalities on joint angular velocities, linearizes the line-of-sight constraints and dynamic equations via quaternion differentiation and first-order Taylor expansion, and ensures the accuracy of convex approximation through successive linearization and trust-region constraints, ultimately converting the original nonlinear non-convex problem into an iteratively solved convex optimization problem.
Finally, the study validates the effectiveness of the proposed method through two scenarios, namely static target monitoring and dynamic target monitoring. Fig. 4 illustrates the motion process of the manipulator in the static target scenario, demonstrating that the manipulator successfully circumvents an obstacle located on its mandatory path during motion. Fig. 5 presents the optimization results for the static scenario, in which subfigures A and B show that the objective function converges to the optimum after 17 iterations, and subfigure L indicates that the end-effector line-of-sight angle remains consistently within the 30° field of view, thereby verifying satisfaction of the line-of-sight constraint. In the dynamic target monitoring scenario, Fig. 6 depicts the process of the manipulator tracking a moving target; the optimization results in Fig. 7 demonstrate that the algorithm converges after 18 iterations, and although the driving torque of joint 3 exhibits brief saturation at certain instants, the end-effector line-of-sight constraint is still effectively satisfied. The solution times for the two scenarios are 112.7 seconds and 122.9 seconds, respectively. This hierarchical planning method effectively overcomes the poor scalability of sampling-based methods under complex constraints, while providing a high-quality initial trajectory for sequential convex programming, offering an effective technical solution for autonomous motion planning of space redundant manipulators under end-effector task constraints such as extravehicular visual monitoring.
