Recoleta
GitHub
Source note

Open-Loop Planning, Closed-Loop Verification: Speculative Verification for VLA

Authors
Zihua Wang, Zhitao Lin, Ruibo Li, Yu Zhang, Xu Yang, Siya Mi, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot ControlAction ChunkingClosed Loop VerificationLibero
Open arXivMarkdown source

Summary

SV-VLA cuts VLA control cost by planning action chunks with a large model and checking them online with a small verifier. The goal is to keep most of the speed of open-loop chunking while recovering some of the adaptability of closed-loop control.

Problem

  • VLA models for manipulation are strong but expensive to run at every control step.
  • Action chunking reduces inference cost by predicting multiple future actions at once, but later actions use stale observations and can drift when the environment changes.
  • Speculative decoding from language modeling does not transfer cleanly to robotics because future action validity depends on future observations that appear only after real execution.

Approach

  • SV-VLA uses the heavy VLA as a low-frequency macro-planner. At a planning boundary, it outputs an action chunk of length K and a planning-context feature from an internal transformer layer.
  • A lightweight verifier runs at each control step on the latest observation, encodes the image with a small vision backbone such as ViT-Tiny, fuses that feature with the saved planning context, and predicts a reference action.
  • The system compares the planned action and the verifier's reference action with a normalized L1 distance. If the distance stays below a threshold (\tau), it executes the planned action.
  • If the distance exceeds (\tau), the system discards the rest of the chunk and calls the heavy VLA again to replan from the current state.
  • The heavy VLA stays frozen during verifier training; only the verifier is trained with an L1 regression loss against ground-truth actions, which keeps the method compatible with pretrained VLAs.

Results

  • The paper reports experiments on the LIBERO benchmark and says SV-VLA improves the average success rate across three subtasks by 11.4 percentage points over the open-loop baseline, from 79.5% to 90.90%.
  • The excerpt does not provide the names of the three subtasks, per-task scores, latency numbers, chunk size used in the main result, or a comparison against step-wise closed-loop VLA.
  • The method claims to preserve the efficiency profile of chunked control: best-case per-step cost approaches (C_{VLA}/K + C_{verify}), while worst-case cost approaches (C_{VLA}) if replanning triggers every step.
  • The text gives one concrete chunking example, noting that long chunks such as (K=64) increase exposure to drift in standard open-loop execution.

Link