Event

EgoForesight: Scaling Policy Learning with Noisy Egocentric Action Data

1 reports · 1 sources

Chubin Zhang and co-authors released a paper proposing EgoForesight, a vision-language-action framework that lets robot demonstration trajectories participate in policy learning together with noisy egocentric videos. Its approach reduces the impact of unreliable action labels through reliability-conditioned noise injection and reliability-weighted regression, and supplements supervision from demonstration outcomes via privileged future representation distillation. The authors report that with 30,000 hours of pretraining data, the method reduces real-world open-loop action MSE by 20.5% and improves RoboCasa few-shot success rate by 8.1 percentage points; after post-training, it also achieves strong absolute performance in simulated and real-world dexterous manipulation.