Vision-Language-Action (VLA) models inherit rich semantic representations from pretrained Vision-Language Models, yet fine-tuning on limited robot demonstrations degrades this structure and undermines generalization. This raises a fundamental question: what constitutes a good action representation?
Inspired by the mirror neuron theory's insight that observation and execution share an intention-level encoding, we examine whether a robot's action representations preserve the semantic structure captured by pretrained encoders. Systematic probing confirms that this structure erodes during fine-tuning, and its quality synchronizes with both task success and out-of-distribution generalization.
To address this, we introduce a plug-and-play method that anchors action representations to a semantic manifold while decomposing them into a shared semantic channel and a private channel, both discarded at inference, leaving the deployed model unchanged.
Validated on different VLA backbones across simulation and real-world benchmarks, our method yields improvements of up to +18.7% on real-world in-distribution tasks and +21.5% on out-of-distribution generalization.
Blogger's Review: This study effectively addresses the generalization issue of robotic action representations through semantic anchoring, providing new insights into the interpretability and performance enhancement of robotic learning, highlighting the potential application of mirror neuron theory in robotics.