CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents
CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents
This paper introduces CaPtain, an open pipeline that converts a small open vision-language model into a capable code-as-policy robot agent. The method uses supervised fine-tuning on synthetic trajectories and multi-turn reinforcement learning with a physical critic to handle delayed credit assignment. Starting from Qwen3.5-4B (0% zero-shot), CaPtain-4B achieves 73.8% success on the LIBERO benchmark after RL, surpassing several frontier-model agents on CaP-Bench and demonstrating sim-to-real transfer on a real robot without target-domain training.
The authors present CaPtain, an open post-training pipeline that transforms a small open-source vision-language model into a capable code-as-policy robot agent without relying on frontier models or pre-trained vision-language-action models at test time.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
improvement on perturbation suites
17.6
View original evidence
ins persist across three perturbation suites (+17.6 to +21.2 points). Without target-domain training, it attains 83.3% on SimplerEnv WidowX, exceeds frontier models on CaP-Bench, and runs on a real robot after simulation-only adaptation. We release all code, data, and model weights.
Open source S6improvement on perturbation suites
21.2
View original evidence
ins persist across three perturbation suites (+17.6 to +21.2 points). Without target-domain training, it attains 83.3% on SimplerEnv WidowX, exceeds frontier models on CaP-Bench, and runs on a real robot after simulation-only adaptation. We release all code, data, and model weights.
Open source S6
- code: not_reported
- paper: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
end-of-episode outcome. We therefore introduce a physical critic that reads privileged simulator state, never seen by the policy, and credits each request by how much it changes the predicted chance of success. Starting from Qwen3.5-4B (0% zero-shot), CaPtain-4B reaches 55.4% held-out LIBERO success after SFT and 73.8% after RL. On LIBERO-PRO, it reaches 78.5%, surpassing several frontier-model agents, and its RL ga
Open source S5
ins persist across three perturbation suites (+17.6 to +21.2 points). Without target-domain training, it attains 83.3% on SimplerEnv WidowX, exceeds frontier models on CaP-Bench, and runs on a real robot after simulation-only adaptation. We release all code, data, and model weights.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org