CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents

Researchers have introduced CaPtain, an open post-training pipeline that transforms small open vision-language models into capable code-as-policy robot agents without requiring a vision-language-action model or frontier model at test time. The system uses supervised fine-tuning followed by multi-turn reinforcement learning guided by a physical critic. Starting from a base model with zero-shot capability, CaPtain-4B achieved significant performance improvements on simulated and real-world manipulation benchmarks. Open article →