Recoleta
GitHub
Source note

VeriSpace: Spatially Grounded Action Verification for Vision-Language-Action Models

Authors
Guiyu Zhao, Longteng Guo, Junyou Zhu, Jun Fu, Yanghong Mei, Bin Cao, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionAction VerificationRobot ManipulationSpatial ReasoningRgb D Perception
Open arXivMarkdown source

Summary

VeriSpace is a test-time action verifier for vision-language-action robot policies. It samples multiple actions, scores them with RGB-D spatial reasoning, and executes the highest-scoring action.

Problem

  • VLA policies often predict one action and execute it immediately, so small pose or gripper errors can cause missed grasps, collisions, or wrong task progress.
  • Existing action verifiers that score candidates from 2D images can miss small 3D differences in contact, clearance, and object alignment.
  • The problem matters because robot actions change the physical scene, and a bad action can make later recovery harder or unsafe.

Approach

  • A frozen VLA policy samples candidate 7D actions, including 6-DoF end-effector motion and gripper command. VeriSpace scores each candidate and selects the top-ranked action.
  • The verifier uses RGB-D input. It back-projects depth into 3D scene coordinates using camera parameters, then encodes those coordinates with sinusoidal features and an MLP.
  • Dual-path scene encoding creates explicit 3D geometry tokens and visual tokens with 3D position injected into CLIP image features.
  • Geometry-guided local aggregation uses farthest point sampling and radius neighborhoods in 3D to capture local shape cues near object surfaces, edges, gaps, and possible contact zones.
  • Training uses a Bradley-Terry pairwise preference loss for action ranking plus cross-entropy supervision for chain-of-thought spatial reasoning. The reported verifier uses LLaVA-7B with LoRA.

Results

  • On SimplerEnv-WidowX with OpenVLA, average success rose from 37.0% to 55.0% across 4 tasks and 50 trials per task, a gain of 18.0 percentage points.
  • On the same OpenVLA setting, VeriSpace beat the strongest listed verifier baseline, Robomonkey, by 14.5 points on average: 55.0% versus 40.5%.
  • OpenVLA task gains were 54.0% to 76.0% on Eggplant in Basket (+22.0), 22.0% to 34.0% on Carrot on Plate (+12.0), 28.0% to 62.0% on Stack Cubes (+34.0), and 44.0% to 48.0% on Spoon on Towel (+4.0).
  • With π0-FAST, average success rose from 57.0% to 60.5%, a gain of 3.5 points. The strongest listed verifier baseline in that block was Robomonkey at 58.5%.
  • The paper states that VeriSpace surpasses the previous best method on SIMPLER by 13.0 points, but the provided table excerpt shows a larger 14.5-point margin over the listed OpenVLA verifier baselines.
  • The excerpt says real-world in-distribution and out-of-distribution experiments improved, but it does not provide the real-world success rates in the visible text.

Link