CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation

Letian Fu, Justin Yu, Karim El-Refai, Ethan Kou, Haoru Xue, Huang Huang, Wenli Xiao, Li Fei-Fei, Guanya Shi, Jiajun Wu, S. Shankar Sastry, Yuke Zhu, Ken Goldberg, Linxi Fan
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:32155-32212, 2026.

Abstract

Code-as-Policy (CaP) is a paradigm in which a language or vision-language model generates executable robot control programs, yet its effectiveness as an autonomous controller for embodied manipulation remains underexplored. Prior CaP systems often rely on high-level, human-designed primitives, making it difficult to separate agent capability from designer-provided scaffolding. We present CaP-X, an open-access framework for systematically studying Code-as-Policy agents in robot manipulation. CaP-X includes four components. CaP-Gym is an interactive environment in which coding agents control robots by synthesizing and executing programs that compose perception and control primitives. Building on this foundation, CaP-Bench evaluates frontier language and vision-language models across varying levels of abstraction, interaction, and perceptual grounding. Across 12 models, the task success rates improve with human-crafted abstractions but degrade as these priors are removed, exposing a dependence on designer scaffolding. At the same time, we observe that scaling test-time computation with multi-turn interaction, structured execution feedback, visual differencing, automatic skill synthesis, and ensembled reasoning can substantially improve robustness even when agents operate over low-level primitives. These findings motivate CaP-Agent0, a training-free framework that achieves near human-level reliability on several manipulation tasks in simulation and on real embodiments. CaP-RL explores reinforcement learning with verifiable rewards to improve success rates and supports sim-to-real transfer through a shared code-as-action-space interface. Together, CaP-X provides an open-access platform for advancing embodied coding agents. Project page: https://capgym.github.io

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-fu26u, title = {{C}a{P}-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation}, author = {Fu, Letian and Yu, Justin and El-Refai, Karim and Kou, Ethan and Xue, Haoru and Huang, Huang and Xiao, Wenli and Fei-Fei, Li and Shi, Guanya and Wu, Jiajun and Sastry, S. Shankar and Zhu, Yuke and Goldberg, Ken and Fan, Linxi}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {32155--32212}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/fu26u/fu26u.pdf}, url = {https://proceedings.mlr.press/v306/fu26u.html}, abstract = {Code-as-Policy (CaP) is a paradigm in which a language or vision-language model generates executable robot control programs, yet its effectiveness as an autonomous controller for embodied manipulation remains underexplored. Prior CaP systems often rely on high-level, human-designed primitives, making it difficult to separate agent capability from designer-provided scaffolding. We present CaP-X, an open-access framework for systematically studying Code-as-Policy agents in robot manipulation. CaP-X includes four components. CaP-Gym is an interactive environment in which coding agents control robots by synthesizing and executing programs that compose perception and control primitives. Building on this foundation, CaP-Bench evaluates frontier language and vision-language models across varying levels of abstraction, interaction, and perceptual grounding. Across 12 models, the task success rates improve with human-crafted abstractions but degrade as these priors are removed, exposing a dependence on designer scaffolding. At the same time, we observe that scaling test-time computation with multi-turn interaction, structured execution feedback, visual differencing, automatic skill synthesis, and ensembled reasoning can substantially improve robustness even when agents operate over low-level primitives. These findings motivate CaP-Agent0, a training-free framework that achieves near human-level reliability on several manipulation tasks in simulation and on real embodiments. CaP-RL explores reinforcement learning with verifiable rewards to improve success rates and supports sim-to-real transfer through a shared code-as-action-space interface. Together, CaP-X provides an open-access platform for advancing embodied coding agents. Project page: https://capgym.github.io} }
Endnote
%0 Conference Paper %T CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation %A Letian Fu %A Justin Yu %A Karim El-Refai %A Ethan Kou %A Haoru Xue %A Huang Huang %A Wenli Xiao %A Li Fei-Fei %A Guanya Shi %A Jiajun Wu %A S. Shankar Sastry %A Yuke Zhu %A Ken Goldberg %A Linxi Fan %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-fu26u %I PMLR %P 32155--32212 %U https://proceedings.mlr.press/v306/fu26u.html %V 306 %X Code-as-Policy (CaP) is a paradigm in which a language or vision-language model generates executable robot control programs, yet its effectiveness as an autonomous controller for embodied manipulation remains underexplored. Prior CaP systems often rely on high-level, human-designed primitives, making it difficult to separate agent capability from designer-provided scaffolding. We present CaP-X, an open-access framework for systematically studying Code-as-Policy agents in robot manipulation. CaP-X includes four components. CaP-Gym is an interactive environment in which coding agents control robots by synthesizing and executing programs that compose perception and control primitives. Building on this foundation, CaP-Bench evaluates frontier language and vision-language models across varying levels of abstraction, interaction, and perceptual grounding. Across 12 models, the task success rates improve with human-crafted abstractions but degrade as these priors are removed, exposing a dependence on designer scaffolding. At the same time, we observe that scaling test-time computation with multi-turn interaction, structured execution feedback, visual differencing, automatic skill synthesis, and ensembled reasoning can substantially improve robustness even when agents operate over low-level primitives. These findings motivate CaP-Agent0, a training-free framework that achieves near human-level reliability on several manipulation tasks in simulation and on real embodiments. CaP-RL explores reinforcement learning with verifiable rewards to improve success rates and supports sim-to-real transfer through a shared code-as-action-space interface. Together, CaP-X provides an open-access platform for advancing embodied coding agents. Project page: https://capgym.github.io
APA
Fu, L., Yu, J., El-Refai, K., Kou, E., Xue, H., Huang, H., Xiao, W., Fei-Fei, L., Shi, G., Wu, J., Sastry, S.S., Zhu, Y., Goldberg, K. & Fan, L.. (2026). CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:32155-32212 Available from https://proceedings.mlr.press/v306/fu26u.html.

Related Material