RL-Driven Data Generation for Robust Vision-Based Dexterous Grasping

2025-04-25

AI Summary

Vision-action models for dexterous grasping suffer from poor generalization and instability when adapting to objects with diverse geometries. Method: This paper proposes a reinforcement learning–driven simulation data generation framework. It features a modular architecture combining parameterized reference trajectories with a PPO-based residual policy, balancing demonstration fidelity and cross-object geometric adaptability; integrates contact-aware trajectory modeling and vision-conditioned policy networks to automatically synthesize high-quality, contact-rich, geometry-diverse grasping trajectories in simulation; and leverages real-to-sim-to-real transfer alongside simulation data augmentation. Contribution/Results: Experiments demonstrate significant improvements in zero-shot grasping success rates on unseen objects, effectively mitigating the training bottleneck caused by scarcity of real-world grasping data.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Imitation Learning & Inverse Reinforcement Learning
• Intelligent Robots: Manipulation

Application Category

• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Web data generation and simulation

Abstract

This work presents reinforcement learning (RL)-driven data augmentation to improve the generalization of vision-action (VA) models for dexterous grasping. While real-to-sim-to-real frameworks, where a few real demonstrations seed large-scale simulated data, have proven effective for VA models, applying them to dexterous settings remains challenging: obtaining stable multi-finger contacts is nontrivial across diverse object shapes. To address this, we leverage RL to generate contact-rich grasping data across varied geometries. In line with the real-to-sim-to-real paradigm, the grasp skill is formulated as a parameterized and tunable reference trajectory refined by a residual policy learned via RL. This modular design enables trajectory-level control that is both consistent with real demonstrations and adaptable to diverse object geometries. A vision-conditioned policy trained on simulation-augmented data demonstrates strong generalization to unseen objects, highlighting the potential of our approach to alleviate the data bottleneck in training VA models.

Problem

• Improving vision-action models for dexterous grasping generalization
• Generating contact-rich grasping data for diverse object shapes
• Alleviating data bottleneck in training vision-action models

Innovation

• RL-driven data augmentation for grasping
• Parameterized reference trajectory refinement
• Vision-conditioned policy for generalization