Procedural Digital Cousins for Scalable Robot Learning in Space

Andrej Orsula, Ricard M Castan, Miguel Olivares-Mendez, Carol Martinez

Reviewer: ~Masafumi_Endo1

Keywords: Space Robotics, Reinforcement Learning, Sim-to-Real Transfer

TL;DR: We use generative digital cousins to train RL policies across procedurally generated space simulations to enable zero-shot deployment of autonomous behaviors.

Abstract: Robots play a transformative role in the future of space exploration, from fleets of autonomous rovers navigating unstructured planetary surfaces to robotic manipulators constructing orbital megastructures. However, the adoption of data-intensive robot learning methods in this domain is severely hindered by extreme data scarcity and the practical infeasibility of physical training upon deployment. While high-fidelity digital twins are highly effective for terrestrial workflows, constructing an exact virtual replica is often infeasible for the unpredictable off-world environments. To address this limitation, we introduce the Space Robotics Bench, an open-source framework that leverages procedural content generation and domain randomization to synthesize a vast distribution of physically plausible scenarios. To ensure reliable sim-to-real transfer, the foundational parameters of these generative environments are anchored to physical analogue facilities via the incorporation of digital cousins across both planetary and orbital operations. We establish baselines using reinforcement learning algorithms and demonstrate the efficacy of the framework by successfully achieving zero-shot sim-to-real transfer of a learned navigation policy. Our results show that agents trained across diverse procedural simulations achieve robustness, which highlights the generative approach as a highly scalable pipeline for providing the statistical assurance required to deploy autonomous systems beyond Earth.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 31