Generalizable robot control beyond VLAs with pretrained Internet-scale video model backbones

A next-generation diffusion-based control architecture built around a newly designed 16‑DoF tendon‑driven humanoid hand. Using glove and VR teleoperation data, MIMIC‑ONE achieves up to 93.3 % task success (even on unseen tasks) by learning smooth, self-correcting motion policies from sensory inputs alone.

A novel diffusion-based framework to align action spaces across human hands, mimic’s humanoid robotic hands, and other grippers. By training a unified latent action space via contrastive learning, LAD enables a single policy to control diverse robot embodiments, delivering up to a ~25 % lift in manipulation success and unlocking scalable skill transfer across robotic platforms

In collaboration with ETH Zürich and Microsoft Research, MAPLE introduces a new method to train real-world dexterous robotic hands by learning from large-scale video of human hand interactions. It shows how manipulation priors extracted from egocentric footage enable robots to grasp, manipulate, and generalize across varied tasks.