Training-free correction boosts VLA robot success rates by 28.8%

Vision-Language-Action (VLA) models have shown impressive generalization in robotic tasks, but they suffer from a critical limitation: most are trained on single-frame observations, rendering them structurally blind to temporal dynamics. When faced with moving objects or changing environments, these models degrade severely. Existing fixes require costly retraining or introduce latency and temporal inconsistency. In a new paper, Zhang et al. introduce Pace-and-Path Correction (PPC), a training-free, closed-form operator that wraps any chunked-action VLA at inference time. From a single quadratic cost, PPC jointly optimizes two orthogonal channels: a pace channel that compresses execution along the planned direction (speeding up or slowing down) and a path channel that applies a spatial offset to adapt to motion. This simple yet principled approach absorbs perceived dynamics within the action chunk window without needing additional training data or fine-tuning.

To evaluate, the team built MoveBench, a diagnostic benchmark isolating motion as the sole controlled variable. Results show PPC consistently outperforms state-of-the-art training-free wrappers and dynamic-adaptive methods. In dynamic-only environments, it boosts absolute success rates by up to 28.8% over foundational VLA models; in mixed static-dynamic settings, the gain reaches 25.9%. The method works across different VLA backbones and action chunk sizes, making it a practical drop-in solution for robotics teams deploying VLAs in real-world, non-stationary conditions. This work highlights that tackling temporal blindness doesn't require heavy infrastructure upgrades—sometimes a clever inference-time correction is enough.