Computational Robotics Group, Harvard

BEACON Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation

Antong Zhang1 Han Qi2 Heng Yang2

1Brown University 2Harvard University

arXiv Code BibTeX

BEACON co-trains a diffusion policy on hundreds of simulated demonstrations and a few dozen from the target domain, and learns how much each source demonstration should count.

How it works

Setup

500 simulated source demonstrations per task; 20–40 target demonstrations in a shifted simulator or on the real robot. Fixed-ratio co-training counts every source trajectory the same, whether or not it resembles the target.

Score each source sample

Every source window gets a discrepancy score di: the distance from its encoder embedding to its k nearest target embeddings. Near 1 means in-distribution; large means mismatched.

Learn weights and policy together

A surrogate of the Best-Effort Adaptation bound: q-weighted imitation loss, a discrepancy penalty, and a regularizer. Alternating updates of the source weights and the diffusion policy; target samples always stay in.

Several sources

MS-BEACON adds a domain-level weight on top of the per-sample weights, so simulators closer to the target get more say.

What BEACON achieves

Stronger sim-to-sim, sim-to-real and multi-source transfer than target-only, fixed-ratio and feature-alignment baselines, and it holds up on out-of-distribution initial states.

Abstract

We introduce BEACON, Best-Effort Adaptation for Cross-Domain Co-Training, a theory-guided framework for training generative robot policies with abundant source demonstrations and limited target demonstrations. BEACON casts cross-domain co-training as a discrepancy-aware importance-reweighting problem, jointly learning a diffusion-based visuomotor policy and per-sample source weights through an objective inspired by the Best-Effort Adaptation (BEA) bound. To make this approach practical for high-dimensional sequence policies, we develop scalable instance-level discrepancy proxies, stochastic alternating updates for policy and weights, and a multi-source extension that balances heterogeneous source domains. Across sim-to-sim, sim-to-real, and multi-source manipulation settings, BEACON improves robustness and data efficiency over target-only, fixed-ratio co-training, and feature-alignment baselines. Without an explicit alignment objective, BEACON also brings source and target representations closer together as an implicit result of discrepancy-aware co-training.

BEACON rollouts

Real-robot trials are recorded by the robot’s camera. OOD trials start from object positions outside the range of the target demonstrations.

Citation

Published at NeurIPS 2026.

@inproceedings{zhang2026beacon,
  title     = {{BEACON}: Cross-Domain Co-Training of Generative Robot Policies
               via Best-Effort Adaptation},
  author    = {Zhang, Antong and Qi, Han and Yang, Heng},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2605.08571}
}