Daxiao Robotics, in collaboration with institutions such as the Chinese University of Hong Kong and Nanyang Technological University, has open-sourced ACE-eGO-Hand, a two-hand 3D shape reconstruction model for first-perspective videos, and has produced about 5,000 hours of training data. The model transforms the video diffusion model into a deterministic geometric encoder, uses a single feedforward instead of multi-step denoising, and speeds up inference by about 33 times. It is reported that this technology can directly transform human-first perspective operation videos into structured 3D trajectories and migrate them to dexterous hands, greatly reducing the data collection threshold for emulated intelligent imitation learning.

Daxiao Robotics, in collaboration with institutions such as the Chinese University of Hong Kong and Nanyang Technological University, has open-sourced ACE-eGO-Hand, a two-hand 3D shape reconstruction model for first-perspective videos, and has produced about 5,000 hours of training data. The model transforms the video diffusion model into a deterministic geometric encoder, uses a single feedforward instead of multi-step denoising, and speeds up inference by about 33 times. It is reported that this technology can directly transform human-first perspective operation videos into structured 3D trajectories and migrate them to dexterous hands, greatly reducing the data collection threshold for emulated intelligent imitation learning.