Announcement
Researchers say Grasping the Future forecasts hand-and-object motion 4.5 seconds ahead
The team says its model learns joint 3D hand-and-object motion from internet videos, fingers included.
TLDR
Half a second in, 20 possible futures out: the team says Grasping the Future forecasts the next 4.5 seconds of hands and objects moving together in 3D. Its training data includes 62,562 interactions reconstructed from internet videos and eight motion-capture datasets, according to the project page. The researchers report a 73.8% success rate when simulated robot hands follow its forecasts using physics-aware retargeting.
Combined views
—
6 Sources, first seen ago

Comments