AstroBench: Catching an object in weightlessness in a space station, an exercise that generalist robotic models almost always miss

An open test bench simulates the capture of floating objects and maintenance tasks inspired by the ISS: a generalistic model captures only 6 targets out of 400.

In a spacecraft, an escaping object continues to rotate and drift, and its trajectory changes as soon as the gripper touches it. A team publishes on arXiv AstroBench, an open testbed dedicated to this problem, with a fixed-base robotic arm. It combines four initial motion scenarios with sample storage operations modeled after the International Space Station's MELFI freezer and unscrewing tasks inspired by NASA's Robotic Refueling Mission: nine tasks and 29 task-target combinations in total, with more than 8,400 objects from Objaverse and five cabin scenery.

Contrasting results

The authors tested existing models without retraining. The general-purpose π0.5 DROID model captured only 2 stationary targets out of 200 and 4 moving targets out of 200. Conversely, an ACT policy trained specifically on cubes caught 8 out of 10 stationary and 6 out of 10 in motion. For the team, these discrepancies show that the direct transfer of generalist models to microgravity remains difficult, and point to motion perception and contact-sensitive control as priority avenues. The code and environments are published on GitHub.