He Wang

He Wang focuses on embodied AI, aiming to equip robots with generalized capabilities that enable them to perform complex tasks in challenging environments. By integrating advanced AI technologies into robotics, his work addresses pressing societal issues such as aging populations and labor shortages.

One of the central challenges in embodied AI is the scarcity and high cost of high-quality training data. To overcome this, He has devoted years to developing a large-scale, high-fidelity synthetic dataset with precise physical realism. This dataset advances the development of generalized skill learning and large embodied models, and has led to breakthroughs in areas such as generalized grasping and end-to-end embodied models.

He introduced DexGraspNet, the first million-scale dataset tailored for dexterous hand grasping, boosting the efficiency of synthetic data generation by 50-fold. Building on this, his team developed the UniDexGrasp++ algorithm, which enables highly reliable generalized grasping across thousands of objects.

He also pioneered GraspVLA, the first end-to-end foundational model for embodied grasping. Trained entirely on synthetic data, comprising one billion vision-language-action (VLA) frame pairs, GraspVLA achieves generalized closed-loop grasping without relying on large-scale real-world data, setting a new benchmark for VLA model development. With only a small amount of fine-tuning, this “generalist” model can quickly adapt to become an “expert” in specific scenarios. This approach offers significant advantages—large-scale data, strong generalization, and low cost—addressing key bottlenecks in the development of embodied intelligence. It is expected to drive large-scale commercialization of end-to-end embodied foundation models as early as 2025.

As the founder and CTO, He Wang co-founded Beijing Galbot Co., Ltd. in May 2023. To date, the company has raised 1.3 billion yuan in funding.