In context
As space missions target larger and modular structures, on-orbit assembly by free-floating space robots becomes critical. However, assembly object vibration and robot-structure dynamic coupling challenge traditional compliant control, prompting research into learning-based methods.
What was reported
Researchers from Northwestern Polytechnical University and the China Academy of Space Technology proposed a model-data hybrid approach combining impedance control with deep reinforcement learning for on-orbit assembly of a segmented space telescope. The task was modeled as a Markov decision process, with joint impedance control serving as a prior model to improve learning efficiency. The proximal policy optimization (PPO) algorithm was used to train the assembly strategy, addressing dynamic coupling and vibration effects.
To accelerate validation, a parallelized training and testing environment was built using Isaac Gym. Simulations demonstrated improved compliant control performance and robustness against uncertainties, including base pose deviations and sensor noise, via domain randomization.
The method avoids the need for precise trajectory specification and contact-state identification required by traditional impedance or hybrid force/position control, offering a more adaptive solution for complex assembly tasks.
Why it mattered
This work highlights the potential of reinforcement learning to enhance compliant control in space robotics, paving the way for autonomous assembly of large space structures. The model-data hybrid approach could also inspire terrestrial industrial automation where dynamic coupling and uncertainty complicate precision assembly.
"A model-data hybrid driving approach is proposed that integrates impedance control with deep reinforcement learning to enable efficient learning of assembly strategies."
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-03-15 · “基于强化学习的空间机器人在轨装配柔顺控制方法”