SynGrasp-1B

A billion-frame simulated synthetic grasping dataset built by Galbot and others, used to pretrain GraspVLA.

SynGrasp-1B is a synthetic grasping dataset built by Galbot (银河通用) together with Peking University, the University of Hong Kong, and the Beijing Academy of Artificial Intelligence, totaling about 1 billion frames, all generated in simulation: 240 categories and more than 10,000 object models were selected from Objaverse and randomly placed on tables, BoDex generated stable grasp poses, cuRobo planned the grasping trajectories, and the results were rendered after domain randomization of materials, lighting, camera viewpoint, background, and initial pose. It's used to pretrain the grasping VLA model GraspVLA (CoRL 2025), testing whether synthetic action data alone can transfer zero-shot to real-world grasping. The dataset was made public on Hugging Face in August 2026.

ExampleGraspVLA, pretrained on SynGrasp-1B with no real-robot fine-tuning at all, can grasp objects it never saw during training on a real tabletop, following a language instruction.

Also called

SynGrasp

Related

GraspVLA · Synthetic Data · Domain Randomization · Objaverse · cuRobo (NVIDIA GPU-accelerated motion planning) · Sim-to-Real Transfer

Sources

GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data (arXiv 2505.03233)
GraspVLA 项目页 (Chinese)
PKU-EPIC/GraspVLA (GitHub)

As of

2026-08