Learning to Detumble: Adaptive Post-Capture Stabilization of Uncooperative Space Debris

Álvaro Belmonte-Baeza, Celia Redondo-Verdú, Jose L. Ramon, Gabriel J. Garcia, Miguel Cazorla, Jorge Pomares

Keywords: Reinforcement Learning, Active Debris Removal, On-Orbit Servicing, Space Robotics

TL;DR: This paper presents an RL framework for a dual-arm satellite with 6D propulsion to autonomously stabilize and efficiently detumble uncooperative space debris.

Abstract: Active debris removal missions face a critical technical challenge in the autonomous post-capture stabilization of non-cooperative, freely tumbling targets. Traditional model-based control methods often struggle during this phase due to dynamic uncertainties and reliance on accurate state estimates. To address this gap, we propose a Reinforcement Learning (RL) framework specifically designed for the post-capture detumbling of space debris using a dual-arm robotic satellite equipped with a 6D propulsion system. By training the RL agent over a randomized distribution of target states and inertial parameters, our approach handles partial observability and state uncertainty without relying on precise analytical models. Furthermore, the reward formulation explicitly encourages energy-efficient actuation, coordinating the redundant robotic arms, base thrusters, and reaction wheels to actively dissipate the target's kinetic energy. Quantitative evaluations demonstrate that the learned policy effectively dampens multi-axis tumbling motions, achieving near-zero residual velocities across a wide spectrum of initial dynamic states.

Submission Number: 4