Dayi Dong, Maulik Bhatt, Aayushi Shrivastava, Lasse Peters, Negar Mehr
University of California, Berkeley
ALTER: Adaptation from Limited demonstrations for Team coordination with Existing-skill Retention.
Our method, ALTER adapts pretrained diffusion policies to coordinate with other robots while retaining their original independent skills. A trainable residual coordination head corrects a frozen base policy using limited collaborative demonstrations and single-agent replay distilled from the base policy itself.
Paper · Project website · Models · Data
Simulation and hardware research code are available on this repository’s main
branch. Checkpoints and data are public at ALTER-models
and ALTER-data.
Original code/checkpoints use Apache-2.0; original datasets use CC BY 4.0.
See LICENSE and NOTICE for source scope. Hardware checkpoints,
recordings and prepared training caches are available under hardware/v1/ in the
same Hub repositories; use the separate hardware guide
and hardware manifest. Physical operation still requires the
private external robot-control package.
Start with installation, simulation artifact downloads, simulation workflows, and hardware preparation. Licensing review and Hugging Face publication describe the release scope and verification. Local smoke tests do not reproduce the paper tables.
The project website is published from docs/. The website and supporting
media are preserved alongside the research code in this repository.
If you find ALTER useful in your research, please cite:
@article{dong2026alter,
title={Residual Denoising Enables Sample-Efficient Multi-Agent Coordination on Demand},
author={Dong, Dayi and Bhatt, Maulik and Shrivastava, Aayushi and Peters, Lasse and Mehr, Negar},
journal={arXiv preprint arXiv:2609.32129},
year={2026},
url={https://arxiv.org/abs/2609.32129}
}