Learning Robust Strategies For In-Space Autonomous Assembly

Abstract

Autonomy plays a key role for more scalable, precise and economic future robotic space missions. Teleoperated space robotic tasks are affected by the communication delay between the spacecraft and the ground station. In the context of robotics in-space assembly and the PULSAR project, a technical demonstrator of the autonomous assembly of a telescope's primary mirror, a learning-based method for such operation is proposed in this work. Conventional robotics assembly methods usually rely on pre-defined motions and strategies, and are engineered for a single use case. Learning-based approaches allow to use the same method for different geometries with little efforts. In this work, a reinforcement learning environment where an industrial robotic arm performs the assembly operation is modelled. Then, open source software components are used to implement the proposed design and validate it in a simulated environment with a precise physics engine. The experiments in the simulator show that the training converges and the trained reinforcement learning agents are able to successfully perform the assemblies of different peg-in-hole geometries and the parts designed for the PULSAR project. These results make reinforcement learning methods worth considering for future real-world experiments and potential in-space assembly missions.