Towards Benchmarking Robotic Manipulation in Space

As humanity ventures deeper into space, the demand for autonomous robotic systems capable of performing complex manipulation sequences is becoming increasingly critical. This work introduces a set of tasks designed to explore learning-based approaches in the context of space robotics while emphasizing the need for generalization and adaptability. The benchmark leverages procedural generation and parallel simulation environments to expose agents to a wide range of scenarios across different domains of space. Preliminary results highlight the challenges posed by procedural variability and underscore the importance of evaluating generalization capabilities in the design of the benchmark.