Repository navigation
CaP-X v0.0.1 — Initial Release
CaP-X is an open-source framework for systematically studying Code-as-Policy agents in robot manipulation.
Components
- CaP-Gym: 39 interactive tasks across Robosuite, LIBERO-PRO, and BEHAVIOR (Isaac Sim)
- CaP-Bench: 8-tier benchmark (S1-S4 single-turn, M1-M4 multi-turn)
- CaP-Agent0: Training-free agentic framework with visual differencing and skill libraries
- CaP-RL: GRPO-based reinforcement learning for coding agents
Highlights
- Interactive web UI with 3D visualization (Viser)
- OpenRouter LLM proxy