SKYWALKER: Dual-Agent Reinforcement Learning for Autonomous Crawling in a Planar Microgravity Analog : Sim-to-Real Evaluation at ESA’s Orbital Robotics Lab

SKYWALKER is a dual-agent reinforcement learning framework that enables a free-floating robotic platform to "crawl" hand-over-hand between fixed anchor points in a planar microgravity analog. The task is divided between two cooperating RL agents, one for arm-based anchoring and one for base repositioning. Both agents are trained in a frictionless simulation environment.In hardware trials conducted at the European Space Agency’s Orbital Robotics Lab (ORL) as part of the ESA Academy Experiments Programme, the learned arm policy (Agent A) successfully performed autonomous docking, while the base-motion policy (Agent B) encountered sim-to-real transfer issues and was temporarily replaced with a scripted fallback. Subsequently, Agent B was refined and validated on a physical ball-bearing testbed at Aalborg University, enabling fully learned anchor-to-anchor crawling.These results demonstrate the feasibility of learning-based control for on-orbit assembly tasks and highlight the critical role of high-fidelity simulation and robust safety measures in deploying reinforcement learning for space robotics.