Lending a hand to Mimic Robotics

Mimic advances dexterous manipulation by leveraging large datasets to improve model generalization for its robotic hand.

We provided them with 30 hours of fully annotated, egocentric human demonstrations of clothing-folding tasks.

Highlights

• Large-scale diversity: Our network delivered a wide range of demonstrations across varied environments and included clothing of different materials, sizes, colors, and textures.

• Low cost: Mimic significantly reduced expenses compared to organizing and collecting the data in-house.

• Time-to-market & flexibility: Data collection was completed in just a few days, and Mimic could easily request additional demonstrations and custom annotations.

The final dataset included over 7,200 demonstrations, gathered by more than 60 data collectors using over 1,200 unique objects. Each demonstration was enriched with language annotations, hand tracking, and bounding boxes, creating a robust dataset for training next-generation Physical AI models.