Event

Semantic Anchoring: Introducing Semantic Anchoring into VLA Action Representations

1 reports · 1 sources

Yuan Xu and other authors propose Semantic Anchoring, a plug-and-play method that anchors the action representations of VLA models to a semantic manifold and splits them into a shared semantic channel and a private channel. The authors claim that this method has been validated on multiple VLA backbones through simulation and real-world benchmarks, with improvements of up to 18.7% on real-world in-distribution tasks and up to 21.5% in out-of-distribution generalization. These representations are all discarded at inference time, and the deployed model itself remains unchanged.