UltraDexGrasp: Learning Universal Dexterous Grasping for Bimanual Robots with Synthetic Data
Sizhe Yang, Yiman Xie, Zhixuan Liang, Yang Tian, Jia Zeng, Dahua Lin, Jiangmiao Pang
AI summary
Problem
Current robotic grasping lacks universal multi-strategy capabilities for bimanual robots, primarily due to a scarcity of high-quality, diverse training data that supports coordinated dual-arm manipulation across varied object shapes, sizes, and weights.
Approach
The authors introduce a data generation pipeline that combines optimization-based grasp synthesis with planning-based demonstration generation to create a large synthetic dataset, then train a transformer-based grasp policy on point cloud inputs to predict control commands for multiple grasp strategies.
Key results
- Curated UltraDexGrasp-20M, a 20-million-frame multi-strategy dataset covering 1,000 objects.
- Achieved 84.0% average success rate in simulation on 600 diverse objects, outperforming baselines by 25.2 percentage points.
- Demonstrated robust zero-shot sim-to-real transfer with an 81.2% real-world success rate across varied objects.
- Open-sourced the data generation pipeline to facilitate future bimanual grasping research.
Why it matters
It provides a scalable, data-driven solution for complex bimanual dexterous manipulation, enabling humanoid and dual-arm robots to handle everyday objects with human-like adaptability.
Abstract
Grasping is a fundamental capability for robots to interact with the physical world. Humans, equipped with two hands, autonomously select appropriate grasp strategies based on the shape, size, and weight of objects, enabling robust grasping and subsequent manipulation. In contrast, current robotic grasping remains limited, particularly in multi- strategy settings. Although substantial efforts have targeted parallel-gripper and single-hand grasping, dexterous grasping for bimanual robots remains underexplored, with data be- ing a primary bottleneck. Achieving physically plausible and geometrically conforming grasps that can withstand external wrenches poses significant challenges. To address these issues, we introduce UltraDexGrasp, a framework for universal dex- terous grasping with bimanual robots. The proposed data- generation pipeline integrates optimization-based grasp syn- thesis with planning-based demonstration generation, yielding high-quality and diverse trajectories across multiple grasp strategies. With this framework, we curate UltraDexGrasp- 20M, a large-scale, multi-strategy grasp dataset comprising 20 million frames across 1,000 objects. Based on UltraDexGrasp- 20M, we further develop a simple yet effective grasp policy that takes point clouds as input, aggregates scene features via unidirectional attention, and predicts control commands. Trained exclusively on synthetic data, the policy achieves robust 2026 IEEE International Conference on Robotics and Automation (ICRA 2026) June 1-5, 2026. Vienna, Austria 979-8-3315-8160-2/26/$31.00 ©2026 IEEE 5526