DexEXO
A Wearability-First Dexterous Exoskeleton for Operator-Agnostic Demonstration and Learning
Engineered an adaptive, pose-tolerant hand exoskeleton that physically aligns human demonstrations with robotic embodiment, enabling highly scalable AI policy training.
-
Adaptive Mechanism Design: Designed a passive slider-based finger interface and a multi-DoF pose-tolerant thumb mechanism. This enables seamless cross-operator data collection by accommodating diverse human hand lengths (140mm to 217mm) without the need for strict joint alignment or per-user calibration.
-
Hardware-Level Embodiment Alignment: Integrated a passive demonstration hand that physically and visually matches the deployed target robot. This hardware-in-the-loop design eliminates visual mismatch and the need for complex image post-processing (e.g., segmentation or inpainting) during data collection.
-
End-to-End AI Integration & Impact: Enabled direct diffusion policy training purely from raw wrist-mounted RGB observations. The system outperformed prior rigid wearables (DexUMI) and traditional teleoperation in user comfort, finger independence, and actual task success rates (e.g., piano playing, scissors cutting).
