Event
BEACON: Cross-Domain Co-Training for Generative Robot Policies
1 reports · 1 sources
Antong Zhang, Han Qi, and Heng Yang propose BEACON, a cross-domain co-training framework that trains diffusion model-based visuomotor policies using abundant source-domain demonstrations and limited target-domain demonstrations, and formulates co-training as a discrepancy-aware importance reweighting problem. The method introduces an instance-level discrepancy proxy, stochastic alternating updates of policy and weights, and a multi-source extension. The authors claim that in sim-to-sim, sim-to-real, and multi-source manipulation settings, BEACON improves robustness and data efficiency compared with target-data-only, fixed-ratio co-training, and feature-alignment baselines.
Report timeline
- BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation
BEACON proposes a theory-guided cross-domain co-training framework that trains diffusion-based visuomotor policies using abundant source demonstrations and limited target demonstrations, and formulates co-training as a discrepancy-aware importance reweighting problem. The method introduces scalable instance-level discrepancy proxies, stochastic alternating updates of policy and weights, and a multi-source extension, improving robustness and data efficiency over target-only, fixed-ratio co-training, and feature-alignment baselines in sim-to-sim, sim-to-real, and multi-source manipulation settings.
arXiv cs.RO ·