Event

JevVLA: An Energy-Based Adaptation Framework for Frozen Foundation Policies

1 reports · 1 sources

Wei Wang and other authors propose JevVLA, a lightweight energy-based adaptation framework for vision-language-action (VLA) policies. Its approach is, while keeping the foundation policy frozen, to learn a differentiable scalar energy function conditioned on observations, task context, and actions, used both for discrete candidate comparison and for continuous action refinement or generation; training signals can come from execution feedback, expert corrections, or demonstration preferences. The authors state that each benchmark uses independently trained weights and fixed input/output encodings, but shares the same core network, with about 80,000 trainable parameters. The implementation will be open-sourced.