Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

V-CAGE: Context-Aware Generation and Verification for Scalable Long-Horizon Embodied Tasks

Yaru Liu, Ao-bo Wang, Nanyang Ye

TL;DR

V-CAGE addresses the challenge of learning long-horizon embodied tasks from synthetic data by identifying geometric inconsistency and semantic misalignment as core failure modes. It introduces a closed-loop framework with three hierarchical modules: Semantic Instruction Grounding via an LLM, Context-Aware Scene Instantiation that maintains a dynamic map of prohibited volumes to prevent collisions, and a VLM-Guided Rejection Sampling loop that serves as a visual critic to enforce semantic correctness. Key contributions include the context-aware generation mechanism, a VLM-based rejection sampling process, and empirical evidence that data generated by V-CAGE markedly improves downstream policy success and generalization. The approach advances scalable synthetic data generation by coupling geometric fidelity with semantic verification, enabling more robust long-horizon manipulation in cluttered settings.

Abstract

Learning long-horizon embodied behaviors from synthetic data remains challenging because generated scenes are often physically implausible, language-driven programs frequently "succeed" without satisfying task semantics, and high-level instructions require grounding into executable action sequences. To address these limitations, we introduce V-CAGE, a closed-loop framework for generating robust, semantically aligned manipulation datasets at scale. First, we propose a context-aware instantiation mechanism that enforces geometric consistency during scene synthesis. By dynamically maintaining a map of prohibited spatial areas as objects are placed, our system prevents interpenetration and ensures reachable, conflict-free configurations in cluttered environments. Second, to bridge the gap between abstract intent and low-level control, we employ a hierarchical instruction decomposition module. This decomposes high-level goals (e.g., "get ready for work") into compositional action primitives, facilitating coherent long-horizon planning. Crucially, we enforce semantic correctness through a VLM-based verification loop. Acting as a visual critic, the VLM performs rigorous rejection sampling after each subtask, filtering out "silent failures" where code executes but fails to achieve the visual goal. Experiments demonstrate that V-CAGE yields datasets with superior physical and semantic fidelity, significantly boosting the success rate and generalization of downstream policies compared to non-verified baselines.

V-CAGE: Context-Aware Generation and Verification for Scalable Long-Horizon Embodied Tasks

TL;DR

V-CAGE addresses the challenge of learning long-horizon embodied tasks from synthetic data by identifying geometric inconsistency and semantic misalignment as core failure modes. It introduces a closed-loop framework with three hierarchical modules: Semantic Instruction Grounding via an LLM, Context-Aware Scene Instantiation that maintains a dynamic map of prohibited volumes to prevent collisions, and a VLM-Guided Rejection Sampling loop that serves as a visual critic to enforce semantic correctness. Key contributions include the context-aware generation mechanism, a VLM-based rejection sampling process, and empirical evidence that data generated by V-CAGE markedly improves downstream policy success and generalization. The approach advances scalable synthetic data generation by coupling geometric fidelity with semantic verification, enabling more robust long-horizon manipulation in cluttered settings.

Abstract

Learning long-horizon embodied behaviors from synthetic data remains challenging because generated scenes are often physically implausible, language-driven programs frequently "succeed" without satisfying task semantics, and high-level instructions require grounding into executable action sequences. To address these limitations, we introduce V-CAGE, a closed-loop framework for generating robust, semantically aligned manipulation datasets at scale. First, we propose a context-aware instantiation mechanism that enforces geometric consistency during scene synthesis. By dynamically maintaining a map of prohibited spatial areas as objects are placed, our system prevents interpenetration and ensures reachable, conflict-free configurations in cluttered environments. Second, to bridge the gap between abstract intent and low-level control, we employ a hierarchical instruction decomposition module. This decomposes high-level goals (e.g., "get ready for work") into compositional action primitives, facilitating coherent long-horizon planning. Crucially, we enforce semantic correctness through a VLM-based verification loop. Acting as a visual critic, the VLM performs rigorous rejection sampling after each subtask, filtering out "silent failures" where code executes but fails to achieve the visual goal. Experiments demonstrate that V-CAGE yields datasets with superior physical and semantic fidelity, significantly boosting the success rate and generalization of downstream policies compared to non-verified baselines.
Paper Structure (20 sections, 3 equations, 3 figures, 2 tables)

This paper contains 20 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. LLM-Generated Robotic Simulation Tasks and Training Data
  4. LLMs for Long-Horizon Robotic Task Planning
  5. VLM-Driven Manipulation and Planning
  6. Scalable Synthetic Data Generation
  7. Method
  8. Problem Formulation
  9. Semantic Instruction Grounding
  10. VLM-Guided Rejection Sampling
  11. Generative Scene and Task Synthesis Pipeline
  12. Experiments
  13. Experimental Setup
  14. Main Results
  15. The Critical Role of VLM Verification
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Overview of the V-CAGE framework. A high-level instruction is decomposed by an LLM into subtasks. During simulation, the VLM Gemini3 gemini3report2025 acts as a visual critic after each step, verifying semantic success based on the post-execution image. If a failure is detected, the trajectory generation is aborted to ensure high-fidelity data.
  • Figure 2: The proposed pipeline for autonomous scene generation and task verification. It bridges high-level semantic planning with low-level physical simulation through visual rearrangement and iterative refinement.
  • Figure 3: Example scenes generated by V-CAGE