Robot Learning Beyond Earth: Enabling Adaptive Autonomy in Space

The growing ambition for a sustainable human presence beyond Earth requires autonomous robotic systems capable of reliable operation in extreme and unpredictable conditions. However, developing such autonomy is hindered by the scarcity of extraterrestrial data, the prohibitive cost of hardware testing, and the critical sim-to-real gap. This thesis confronts these obstacles by challenging the conventional pursuit of a singular high-fidelity digital twin. Instead, it proposes a paradigm of diversity over fidelity, where true robotic robustness is achieved not by perfecting one simulation, but by learning to master a massive distribution of scenarios.

To enable such a vision, this work introduces the Space Robotics Bench, a comprehensive open-source simulation framework for robot learning in space that combines scalable parallelization with an integrated procedural engine for the on-demand generation of diverse mission-relevant applications. Building on this foundation, a model-based reinforcement learning methodology is leveraged to acquire robust control policies that can adapt to novel situations.

Experimental validation demonstrates that the principle of procedural diversity yields policies capable of mastering a wide range of mission-critical capabilities, extending from planetary landing and resilient traversal on unstructured deformable terrains to high-precision assembly and tool-aware manipulation. These efforts culminate in the successful zero-shot sim-to-real transfer of a learned policy to a physical rover.

Ultimately, this thesis delivers a new paradigm for the development and validation of learning-based autonomy. By contributing a powerful open-source toolkit and a validated methodological blueprint, this work establishes a scalable pathway for developing and verifying the adaptive robotic systems that will be essential for our multiplanetary future.