EvoScene-VLA: Evolving Scene Beliefs Inside the Action Decoder for Chunked Robot Control
The episode discusses EvoScene-VLA, a method for chunked robot control that allows a robot to maintain and evolve a scene belief by updating it with its own actions. The core idea is embedding this scene belief inside the action decoder, creating a loop where the model corrects its perception against new observations. This design enables persistent, action-updated scene understanding without needing an external world model at inference.