DFM-VLA: Iterative Action Refinement for Robot Manipulation via Discrete Flow Matching

A research team led by Jiayi Chen has introduced DFM-VLA, a novel Vision-Language-Action model that fundamentally changes how robots plan and execute tasks. Unlike current models that generate fixed action sequences, DFM-VLA employs a technique called discrete flow matching to create a token-level probability velocity field. This allows the model to iteratively refine the entire action sequence across multiple steps, dynamically correcting errors in early tokens that would otherwise doom the task. The framework uses a two-stage decoding process with an iterative refinement stage followed by deterministic validation for stable convergence.

Extensive testing shows DFM-VLA's iterative approach delivers superior performance. On the challenging CALVIN benchmark, which measures long-horizon manipulation, DFM-VLA achieved an average success length of 4.44. More impressively, it reached a 95.7% average success rate on the LIBERO benchmark, significantly outperforming strong autoregressive, discrete diffusion, and continuous diffusion baselines. The researchers investigated two methods to construct the velocity field—an auxiliary velocity-head formulation and an action-embedding-guided formulation—while retaining high inference efficiency. This breakthrough addresses a core limitation in robotic AI where early, uncorrectable mistakes lead to task failure, paving the way for more reliable and adaptable robots in real-world settings.