Adapting Robot Models with Human Correction Data

micoolcho · x · 2026-07-16

This discussion highlights that while modern vision-language-action (VLA) and world-action models perform well in manipulation, they still struggle to reliably adapt when **transferred to new robots and tasks**. The proposed approach uses **DAgger-style online imitation learning**: - Deploy the robot first - Continuously collect human correction data - Update the policy dynamically, rather than relying solely on offline fine-tuning The author adds that if similar trajectories or human corrections are available, it would be valuable to scale up the accumulation of **DAgger-style data** across more tasks and environments.

Related event: FlowDAgger: Adapting Robot Models via Human Corrections(2 posts)→

Original post →

More from Embodied

Embodied channel →