Event
VLA single-step action generation: biasing toward the noise-end timestep yields a strong policy
1 reports · 1 sources
Yitong Chen et al. released a paper finding that action generation in vision-language-action models (VLA) differs from text-to-image generation: merely biasing the Flow Matching training timestep distribution toward the noise endpoint yields a strong one-step policy on LIBERO, LIBERO-Plus, LIBERO-Pro, and real-robot tasks. The paper states that the irreducible velocity loss of standard Flow Matching decreases significantly toward the noise endpoint on VLA, whereas image generation models typically show a U-shaped distribution. For open-source VLA models of different architectures, directly replacing iterative decoding with a single forward pass often performs comparably to the original multi-step decoding on LIBERO, CALVIN, and SimplerEnv. The paper has released code and training configurations.
Report timeline
- Let It Be Simple: Research on One-Step Action Generation for Vision-Language-Action Models
The paper finds that VLA action generation differs from text-to-image generation: merely biasing the training timestep distribution toward the noise endpoint yields strong one-step policies on LIBERO, LIBERO-Plus, LIBERO-Pro, and real-world robot tasks. For open-source VLA models with different architectures, directly replacing iterative decoding with a single forward pass often performs comparably to the original multi-step decoding on LIBERO, CALVIN, and SimplerEnv.
arXiv cs.RO ·