CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents
CaPtain lifts a 4-billion-parameter robot agent from 0% zero-shot success to 73.8% on held-out LIBERO tasks.
The pipeline first trains Qwen3.5-4B on synthetic trajectories generated with GPT-6 Astra, reaching 55.4%, then adds multi-turn reinforcement learning. A training-only critic uses simulator state to assign credit to robot requests whose effects may emerge several turns later. At test time, the agent writes code against robot APIs without a VLA, frontier model, or privileged state. The authors also report real-robot operation after simulation-only adaptation and release code, data, and model weights.
HF Daily Papers' note
The pipeline first trains Qwen3.5-4B on synthetic trajectories generated with GPT-6 Astra, reaching 55.4%, then adds multi-turn reinforcement learning. A training-only critic uses simulator state to assign credit to robot requests whose effects may emerge several turns later. At test time, the agent writes code against robot APIs without a VLA, frontier model, or privileged state. The authors also report real-robot operation after simulation-only adaptation and release code, data, and model weights.
HF Daily Papers' note
score 5