Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866912891498659840 |
|---|---|
| author | Cui, Zichen Jeff Rayyan, Omar Etukuru, Haritheja Tan, Bowen Andrianarivo, Zavier Teng, Zicheng Zhou, Yihang Mehta, Krish Wojno, Nicholas Wu, Kevin Yuanbo Anjaria, Manan H Wu, Ziyuan Mao, Manrong Zhang, Guangxun Shah, Binit Kim, Yejin Chintala, Soumith Pinto, Lerrel Shafiullah, Nur Muhammad Mahi |
| author_facet | Cui, Zichen Jeff Rayyan, Omar Etukuru, Haritheja Tan, Bowen Andrianarivo, Zavier Teng, Zicheng Zhou, Yihang Mehta, Krish Wojno, Nicholas Wu, Kevin Yuanbo Anjaria, Manan H Wu, Ziyuan Mao, Manrong Zhang, Guangxun Shah, Binit Kim, Yejin Chintala, Soumith Pinto, Lerrel Shafiullah, Nur Muhammad Mahi |
| contents | The prevalent paradigm in robot learning attempts to generalize across environments, embodiments, and tasks with language prompts at runtime. A fundamental tension limits this approach: language is often too abstract to guide the concrete physical understanding required for robust manipulation. In this work, we introduce Contact-Anchored Policies (CAP), which replace language conditioning with points of physical contact in space. Simultaneously, we structure CAP as a library of modular utility models rather than a monolithic generalist policy. This factorization allows us to implement a real-to-sim iteration cycle: we build EgoGym, a lightweight simulation benchmark, to rapidly identify failure modes and refine our models and datasets prior to real-world deployment. We show that by conditioning on contact and iterating via simulation, CAP generalizes to novel environments and embodiments out of the box on three fundamental manipulation skills while using only 23 hours of demonstration data, and outperforms large, state-of-the-art VLAs in zero-shot evaluations by 56%. All model checkpoints, codebase, hardware, simulation, and datasets will be open-sourced. Project page: https://cap-policy.github.io/ |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_09017 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models Cui, Zichen Jeff Rayyan, Omar Etukuru, Haritheja Tan, Bowen Andrianarivo, Zavier Teng, Zicheng Zhou, Yihang Mehta, Krish Wojno, Nicholas Wu, Kevin Yuanbo Anjaria, Manan H Wu, Ziyuan Mao, Manrong Zhang, Guangxun Shah, Binit Kim, Yejin Chintala, Soumith Pinto, Lerrel Shafiullah, Nur Muhammad Mahi Robotics Machine Learning The prevalent paradigm in robot learning attempts to generalize across environments, embodiments, and tasks with language prompts at runtime. A fundamental tension limits this approach: language is often too abstract to guide the concrete physical understanding required for robust manipulation. In this work, we introduce Contact-Anchored Policies (CAP), which replace language conditioning with points of physical contact in space. Simultaneously, we structure CAP as a library of modular utility models rather than a monolithic generalist policy. This factorization allows us to implement a real-to-sim iteration cycle: we build EgoGym, a lightweight simulation benchmark, to rapidly identify failure modes and refine our models and datasets prior to real-world deployment. We show that by conditioning on contact and iterating via simulation, CAP generalizes to novel environments and embodiments out of the box on three fundamental manipulation skills while using only 23 hours of demonstration data, and outperforms large, state-of-the-art VLAs in zero-shot evaluations by 56%. All model checkpoints, codebase, hardware, simulation, and datasets will be open-sourced. Project page: https://cap-policy.github.io/ |
| title | Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models |
| topic | Robotics Machine Learning |
| url | https://arxiv.org/abs/2602.09017 |