Introducing CaPtain: A New Approach to Robot Manipulation Using Code-as-Policy
CaPtain transforms small open vision-language models into effective robot manipulation agents.
Recent advancements in robotics have led to the development of CaPtain, a novel post-training pipeline that enables small open vision-language models (VLMs) to function as capable code-as-policy (CaP) agents. This approach eliminates the need for pre-trained vision-language-action models or privileged states during testing.
Training Methodology
CaPtain employs supervised fine-tuning (SFT) on synthetic trajectories generated by GPT-6 Astra, followed by multi-turn reinforcement learning (RL). A unique physical critic is introduced to evaluate agent requests based on their impact on predicted success rates.
Performance Metrics
Starting from the Qwen3.5-4B model, CaPtain-4B achieved a 55.4% success rate on held-out LIBERO tasks after SFT, improving to 73.8% post-RL. It also surpassed several frontier-model agents on LIBERO-PRO with a success rate of 78.5% and demonstrated strong performance across various benchmarks.