AC-VLA: Compositional Learning Boosts Out-of-Distribution Generalization for Robotic VLA Models
Researchers introduce AC-VLA, a plug-and-play framework designed to improve out-of-distribution (OOD) generalization in Vision-Language-Action (VLA) models for robotics. The approach addresses two key failure modes—trajectory overfitting and perceptual shortcuts—by combining an LLM-driven instruction decomposer with a state-conditioned masking strategy. AC-VLA achieves approximately 28% absolute improvement on compositional OOD tasks in benchmark tests, while maintaining strong in-distribution performance.
Why it matters: Improving OOD generalization in VLA models is crucial for enabling robots to reliably perform novel task combinations in real-world environments.
Full story at: arXiv Robotics ↗