GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data

2025-05-06

AI Summary

High acquisition cost and dense annotation requirements of real-world robotic manipulation data severely hinder zero-shot generalization and scalability of vision-language-action (VLA) models. Method: We propose the first VLA foundation model pretrained exclusively on SynGrasp-1B—a billion-scale synthetic grasping dataset—introducing a fully synthetic-data-driven architecture. Our approach integrates chain-of-thought training that jointly models autoregressive perception and flow-matching-based action generation, coupled with domain-randomized rendering and internet-scale semantic alignment to enable efficient sim-to-real transfer. Contribution/Results: The model achieves state-of-the-art performance on both real-world and simulated benchmarks, enabling open-vocabulary, zero-shot grasping generalization and few-shot adaptation to human preferences. To foster reproducibility and community advancement, we publicly release the code, the SynGrasp-1B dataset, and the pretrained model weights.

Technology Category

• Computer Vision: Large Vision Models
• Intelligent Robots: Manipulation
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Embodied foundation models are gaining increasing attention for their zero-shot generalization, scalability, and adaptability to new tasks through few-shot post-training. However, existing models rely heavily on real-world data, which is costly and labor-intensive to collect. Synthetic data offers a cost-effective alternative, yet its potential remains largely underexplored. To bridge this gap, we explore the feasibility of training Vision-Language-Action models entirely with large-scale synthetic action data. We curate SynGrasp-1B, a billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and extensive domain randomization. Building on this, we present GraspVLA, a VLA model pretrained on large-scale synthetic action data as a foundational model for grasping tasks. GraspVLA integrates autoregressive perception tasks and flow-matching-based action generation into a unified Chain-of-Thought process, enabling joint training on synthetic action data and Internet semantics data. This design helps mitigate sim-to-real gaps and facilitates the transfer of learned actions to a broader range of Internet-covered objects, achieving open-vocabulary generalization in grasping. Extensive evaluations across real-world and simulation benchmarks demonstrate GraspVLA's advanced zero-shot generalizability and few-shot adaptability to specific human preferences. We will release SynGrasp-1B dataset and pre-trained weights to benefit the community.

Problem

• Exploring synthetic data for Vision-Language-Action model training
• Bridging sim-to-real gap in robotic grasping tasks
• Achieving open-vocabulary generalization with synthetic action data

Innovation

• Pre-trained on billion-scale synthetic grasping data
• Unified Chain-of-Thought for perception and action
• Mitigates sim-to-real gap via flow-matching