Space Robotics Bench

Space Robotics Bench: many simulated rovers in parallel on Martian and lunar terrain

Space Robotics Bench (SRB) is an open-source collection of simulation environments and tasks for robotics research in space, from orbit to the surfaces of the Moon and Mars. Built on NVIDIA Isaac Sim, it trains and evaluates policies across many procedurally varied worlds instead of one.

Six Space Robotics Bench tasks: a robot arm capturing a satellite above Earth and a satellite approached in orbit, a rover excavating regolith and an arm assembling a solar panel on the Moon, and an arm screwdriving and collecting a sample tube on Mars
Tasks in orbit, on the Moon and on Mars, one column each: debris capture and mobile debris capture, excavation and solar-panel assembly, screwdriving and sample collection.

Features

  • Parallel simulation: many environment instances run in parallel to accelerate learning and evaluation.
  • Procedural generation: terrains, rocks and other assets are generated on demand with SimForge.
  • Domain randomization: extensive randomization of the simulated scenes for robustness and generalization.
  • Gymnasium API: environments plug into standard reinforcement-learning frameworks.
  • ROS 2 interface: interoperates with the ROS 2 and Space ROS ecosystems.
  • Tasks: manipulation, mobile robotics and mobile manipulation, each with interchangeable robots and space domains.

Status

The latest release is Space Robotics Bench 0.1.0 (October 2026), dual-licensed under MIT or Apache 2.0. Its own assets and those generated with SimForge are released under CC0 1.0.