dynamics-blindness

Proposes a training-free inference-time method for Vision-Language-Action models to correct pace and path dynamics, improving success rates by up to 28.8% in dynamic environments.