TAVISVideos
Policy rollouts (third-person view, half speed)
Rollouts of the multi-task π0 policy with a head camera on GR1T2: a third-person view (left) next to the robot's head camera (right). The captions mark the head fixation and the grasp as the GALT detector finds them.
Demonstrations: TAVIS-Head (first-person)
The head-camera stream of teleoperated demonstrations from the released datasets, with the output of the GALT detector overlaid: the episode, the measured lead and the grasping arm. The frame turns green once the head has reached its final fixation, and ticks along the top edge mark the current time (white), the head arrival (green) and the grasp (red).
conditional-pick · conditional information gathering
wait-then-act · temporal monitoring
clutter-pick-cube · clutter disambiguation, target defined by appearance
clutter-pick-lift · clutter disambiguation, target defined by language
multi-shelf-scan · vertical workspace search
Demonstrations: TAVIS-Hands (first-person)
Head camera (top), occluded or masked by design, and the two wrist cameras (bottom), which provide the views needed to find and grasp the target.