CaP-X, a new coding agent suite for robotics. an open-access framework for systematically studying Code-as-Policy agents in robot manipulation. An interactive environment in which agents control robots by synthesizing and executing programs that compose perception and control primitives.
- CaP-Agent0: Agentic harness enabling frontier LM to match or exceed human expert on 4 out of 7 robotics tasks
- CaP-Gym: Connecting coding agents to 187 robotics simulation benchmark tasks widely adopted in the robot learning community
- CaP-RL: Sign-of-life for post-training coding model on robotics tasks: 7B OSS model: 20% -> 72%.
- CaP-Bench: Tracking frontier LM performance on robotics. Current best: 32.3%; Human expert: 88.5%
All code open-sourced!
Project: https://capgym.github.io/
Paper: https://arxiv.org/abs/2603.22435
