September 22, 2026
AI agents are now being tested in robotics: RLE-Bench brings together 48 tasks
RLE-Bench contains 48 tasks for AI agents that design and test robots. The suite covers nine task families: from robot control to mechanical design.
Earlier, similar benchmarks more often assessed whether an agent could train a controller. RLE-Bench tests the full engineering chain: the agent writes and runs code in simulation, reviews the result, and refines the solution.
First run. The suite installs on Linux with Git, Make, uv, and Docker. Task task08 does not require a GPU: the authors suggest first running `rlebench prepare task08`, then running the agent on Mobile Base Design.
Testing it in practice. Task08 gives the agent 2 hours on 4 CPU to design a base and controller for Panda, UR5e, and xArm7. In the example, GPT-6 Astra earned credit for shelf access and payload, but failed the static stability, lateral shift, and rotation checks.
The suite is open under MIT: other teams will be able to replicate runs on the same 48 tasks.
Source
