Event

Authors propose π-StepNFT online reinforcement learning framework for streaming VLA

1 reports · 1 sources

Siting Wang and co-authors propose π-StepNFT, an online reinforcement learning framework for streaming vision-language-action models (VLA). The approach treats transitions in SDE sampling as supervision units and uses variance-normalized transition errors to distinguish positive and negative velocity branches, thereby avoiding the need to compute the final action likelihood. The authors report that, on LIBERO, starting from a shared few-shot SFT initialization, π-StepNFT raises the average success rate of π₀ from 57.6% to 92.0% and that of π₀.₅ from 77.1% to 94.0%; on five test-phase-only visual perturbations in ManiSkill, its average success rate is the highest among the evaluated reinforcement learning baselines. These are the authors' self-reported results.