Recoleta
GitHub
Source note

Pre-VLA: Preemptive Runtime Verification for Reliable Vision-Language-Action and World-Model Rollouts

Authors
Zhen Sun, Yongjian Guo, Haoran Sun, Luqiao Wang, Wei Lu, Jiachi Ji, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionWorld ModelsRuntime VerificationRobot SafetyLibero
Open arXivMarkdown source

Summary

Pre-VLA checks VLA action chunks before robot execution or world-model rollout. On LIBERO, it raises RynnVLA-002 average success from 30.79% to 37.62% with 183.9 ms verification per action chunk.

Problem

  • VLA policies can generate unsafe or low-value action chunks under distribution shift, error buildup, or poor confidence calibration.
  • Bad actions can cause collisions, object drops, kinematic violations, and failed long-horizon manipulation episodes.
  • The same bad actions can waste world-model rendering compute and produce misleading rollouts with drift, blurry frames, distorted relations, or false success.

Approach

  • Pre-VLA encodes the language instruction, visual observation, proprioceptive state, and candidate action chunk with a frozen WorldVLA-style multimodal backbone built on Chameleon with action and state tokenizers.
  • A modality-aware pooling layer separates text, image, state, and action tokens, mean-pools each group, and concatenates the four vectors.
  • A lightweight dual-branch head predicts a binary safety confidence and a continuous critic-derived advantage score for each action chunk.
  • Training uses PPO critic signals, K-step advantage estimation, failure backtracking penalties, task-level advantage normalization, Focal Loss, MSE advantage regression, and soft-threshold calibration.
  • At runtime, a scheduler filters actions before execution or world-model rollout, resamples rejected candidates up to a budget, then uses the highest predicted advantage as fallback.

Results

  • On an independent LIBERO test set, Pre-VLA reports F1 = 0.8303 and accuracy = 0.9542 for action validity discrimination.
  • It reduces the false pass rate for invalid actions to 0.0200 on the same LIBERO test setting.
  • Across four LIBERO suites, it improves average closed-loop success over RynnVLA-002 from 30.79% to 37.62%, a gain of 6.83 percentage points.
  • Average forward verification time is 183.9 ms per candidate action chunk; the paper also states a 100–200 ms pure forward range.
  • The excerpt says task execution steps are reduced, but it does not provide step-count numbers.
  • The excerpt claims fewer misleading world-model rollouts, including less target drift and fewer visual artifacts, but it gives no quantitative world-model metric.

Link