Where Success Breaks: Failure-Boundary Learning for Robust Vision-Language-Action Models

This research addresses the structural asymmetry in Vision-Language-Action (VLA) models trained via supervised fine-tuning (SFT), where expert demonstrations only teach success behavior but not failure boundaries. The paper proposes DLS, a method for Failure-Boundary Learning that discovers, localizes, and shapes the boundary between recoverable deviations and task failure using real-grounded behavioral priors and simulated co-training. DLS improves robustness over SFT and online RL baselines, especially in randomized initial states and unseen visual conditions.