Building Embodied Foundation Models for General-Purpose Robots with He Wang, Founder & CTO at Galbot

In this talk

1. 1:53

   Synthetic data to address the VLA data gap

   Wang says vision-language-action models benefit from pretrained vision and language data, but often perform poorly in unfamiliar settings. He argues that teleoperation cannot supply the scale needed for general-purpose models: public datasets have reached about one million trajectories, while he estimates that far more may be required. Galbot’s pipeline builds interactive simulated scenes, generates candidate actions through methods such as human-video animation and reinforcement learning, then checks and renders valid trajectories.

2. 6:37

   Grasping results from large simulated datasets

   Galbot trained a grasping model with one billion simulated action examples and more than 100 million internet grounding boxes. Wang reports about 82% success on LIBERO without fine-tuning, compared with roughly 60% for π₀ and OpenVLA after fine-tuning on that benchmark’s training split. In a real-world bottle-picking task, 200 demonstrations collected in four hours raised success above 90%; the model also transferred to a different bottle brand and changed arrangements.

3. 11:28

   Extending the approach across tasks and deployments

   Wang describes grasping performance as improving sublinearly with more simulated data, while the gap between simulation and real-world results shrinks at billion-frame scale. Galbot applies the pipeline to grocery shelves, where training data is 99% synthetic, and to pharmacy warehouses. He says the company operates stores in about 10 Chinese cities and more than 30 pharmacy warehouses, with robots running up to six hours per charge and mechanical failure intervals exceeding a month.

4. 16:01

   Dexterous hands, navigation, and learned control

   Galbot has expanded synthetic datasets to garment folding, dexterous grasping, and navigation, including NavForm, which Wang says can run across wheeled robots, legged robots, and drones. For whole-body motion, the company turns motion-capture data into simulated reinforcement learning and then real-world deployment on a two-day pipeline. Wang also describes DEX-NDM, which uses a small amount of real data to learn a differentiable model of hand dynamics and improve transfer of simulated dexterous policies.

5. 24:45

   Synthetic data as a cold-start strategy

   Wang frames embodied AI as a cold-start problem and sees synthetic data as the initial source of scale, with real-world data supplementing it. In Q&A, he contrasts current object-focused robot simulation with autonomous driving, where simulating intelligent pedestrians and vehicles is a major challenge. For longer tasks such as folding garments, he says key points on simulated clothing can guide pick-and-place steps, while more complex tasks still need better ways to generate and validate trajectories.