Daxiao Robotics, with researchers from The Chinese University of Hong Kong and
Nanyang Technological University, open-sourced ACE-Ego-Hand, a first-person
video bimanual 3D hand-shape reconstruction model and released roughly 5,000
hours of training data. The model repurposes a video diffusion architecture into
a deterministic geometric encoder so a single forward pass replaces multi-step
denoising, delivering about a 33x inference speedup. The technology converts
egocentric manipulation video into structured 3D trajectories transferable to
dexterous robot hands, significantly lowering the data-collection threshold for
embodied imitation learning.
