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

THE PROBLEM

This paper focuses on vlaModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.. VeriSpace is a test-time actionCore ConceptsActionA command the robot sends to its motors, controller, or low-level system. verifier for Vision-Language-ActionModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. (VLAModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.) models in robotic manipulationManipulation & TasksManipulationUsing a robot arm or hand to move or interact with objects.. It addresses the brittleness of one-shot actionCore ConceptsActionA command the robot sends to its motors, controller, or low-level system. prediction by enabling multiple candidate actionCore ConceptsActionA command the robot sends to its motors, controller, or low-level system. proposals to be evaluated before executionCore ConceptsExecutionActually carrying out planned or predicted actions on the robot.. The system combines 3D-aware scene encoding (preserving both visual semantics and geometry) with spatially-grounded actionCore ConceptsActionA command the robot sends to its motors, controller, or low-level system. reasoning (evaluating taskCore ConceptsTaskThe job the robot is supposed to complete, such as pick-and-place, navigation, or drawer opening. relevance, geometric validity, and goalCore ConceptsGoalThe desired outcome or target state for a robot task. progress). Experiments show consistent improvements over baselineEvaluation & ResearchBaselineA reference method used for comparison. VLAModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. policies and prior verification methods on both in-distribution and out-of-distributionData, Distributions & Training IssuesOOD (Out-of-distribution)A test situation unlike the data seen during training. robotic tasks. Read the paper by tracking the taskCore ConceptsTaskThe job the robot is supposed to complete, such as pick-and-place, navigation, or drawer opening. definition, the robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. or data assumptions, and the evidence that supports the claimed improvement.

HOW IT WORKS

1

Task framing

The paper frames the work as vlaModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions.. Start here because it defines what success means and which assumptions the rest of the method inherits.

Show more ↓

Figure 1. Towards Reliable Test-Time Action Verification. (A) Instead of directly executing a single predicted action, VeriSpace adopts a test-time generate-and-verify paradigm for robust action selection. (B) Reliable action verification in robotic manipulation requires both fine-grained spatial reasoning and goal-progress assessment, calling for 3D-aware perception and explicit spatial reasoning. (C) VeriSpace delivers consistent gains across public benchmarks, VLA backbones, and real-world tasks.

Figure 2. Overview of VeriSpace. At test time, a frozen VLA policy first samples multiple candidate actions from the current observation and instruction. VeriSpace then evaluates each candidate through two key components: a 3D-aware scene encoding module that injects explicit scene geometry into the visual representation, and a spatial reasoning module that assesses each action under the current scene and task goal. The resulting verification scores are used to rank the sampled actions and select the final action for execution.

Figure 4. Real-world setup and qualitative results. Left: the robotic platform, including a Franka arm equipped with wrist-mounted and external RGB-D cameras. Right: example task executions in real-world and the SimplerEnv benchmark.

Figure 8. Visualization results of task-relevant objects grounding.

2

Core method

VLAModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. models often fail with single wrong actions in manipulationManipulation & TasksManipulationUsing a robot arm or hand to move or interact with objects.. VeriSpace adds a verification layer that evaluates multiple candidate actions using 3D scene geometry and spatial reasoning to pick the best one before executionCore ConceptsExecutionActually carrying out planned or predicted actions on the robot.—boosting success rates on real robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. tasks without retraining the base policyCore ConceptsPolicyThe rule or model that maps observations or states to actions.. When reading the method section, identify the inputs, the learned or engineered representation, and the actionCore ConceptsActionA command the robot sends to its motors, controller, or low-level system. or prediction produced by the system.

Show more ↓

3

Data and supervision

For robotics work, the data story is part of the method: check whether the system depends on teleoperationImitation & Reinforcement LearningTeleoperation (teleop)A human remotely controlling the robot, often to collect demonstrations., simulationSimulation & Sim-to-RealSimulationA virtual environment where robots can be trained or tested., internet video, human labels, or robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. rollouts.

Show more ↓

Figure 6. Our pipeline for generating synthetic embodied chain-of-thought data.

Figure 9. Examples of CoT data annotation.

4

Evaluation evidence

The paper should be judged through its evaluationSimulation & Sim-to-RealEvaluationMeasuring how well a robot system performs. protocol: what data is used, what robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. or simulator is tested, and which baselineEvaluation & ResearchBaselineA reference method used for comparison. comparisons support the claim. Look for the gap between the headline result and the deploymentSimulation & Sim-to-RealDeploymentPutting the trained system on a real robot. setting you would actually care about.

Show more ↓

Figure 5. Success rate and inference runtime as the number of action candidates k increases.

Figure 7. Visualization results of depth estimation.

Figure 11. The scene setting of the real-world experiments.

FIGURES

Figure 3. Spatially-grounded action reasoning process.

KEY RESULTS

WHY DEVELOPERS SHOULD CARE

LIMITATIONS

The main limitation to check is whether the claimed behavior holds outside the paper's reported setup. That means testing across different robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. embodiments, scenes, objects, and data distributions.

WHAT COMES NEXT

The practical next step is independent reproduction with clear baselines, ablations, and stress tests. For a developer, the useful follow-up is to map the paper's vlaModern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. assumptions onto a concrete robotCore ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. stack, then test the smallest version of the method that could run end to end.

RELATED PAPERS