Capturing Dexterity: Large-Scale Demonstration Acquisition and Dynamic Contact Modeling for Bimanual Hand–Object Interaction

Achieving dexterous manipulation in AI and robotics depends on learning from high-quality, generalizable demonstrations with detailed annotations. In this talk, I first present Gigahands, a massive and richly annotated dataset capturing diverse bimanual hand object interactions. Existing datasets often fall short in scale, diversity, and detail, limiting the effective training of generative models and policy learners. Gigahands offers the possibility of overcoming these limitations by providing a robust foundation of diverse demonstrations paired with comprehensive text annotations, which are essential for understanding nuanced manual dexterity. Next, I will introduce a method for dynamic contact capturing for bimanual hand–object interactions. Lastly, I'll introduce some open problems in this direction.