A Knowledge-Driven Generation Method of Legged Control Strategy for Space Climbing Robots

The space adhesive climbing robot can attach to the outer surface of spacecraft and autonomously complete extravehicular inspection and operation tasks, which is an important way to realize long-term unmanned spacecraft service in orbit. Aiming at the problem of insufficient control strategy generalization of the adhesive climbing robot after unexpected changes in spacecraft surface properties, we design a knowledge-driven adhesive climbing reward function under the reinforcement learning framework, and build a knowledge-driven legged control strategy generation framework by combining with the Proximal Policy Optimization-clip algorithm. The results demonstrate that under the combined influence of the “knowledge derivation” reward mechanism and legged movement knowledge from adhesive climbing on flat surfaces, the control strategy performance improves by 200%. The legged control strategy enables the robot to adapt to complex surfaces with slope variations ranging from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$-70^{\circ}$</tex> to 70°, height variations between -100 mm and 100 mm, and gap widths of 100 mm to 180 mm, and the strategy exhibits generalization capabilities when encountering surface characteristic changes that differ from the training environment.