EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras
A single active stereo camera beat passive stereo and held up when wrist views were blocked.
EyeRobot 2.0 swivels two eye viewpoints toward a 3D fixation point, then spends more visual computation near the image centers. The system learns gaze control and fixation selection from real-world data, paired with a behavior-cloned gripper policy. In the reported trials, passive stereo fell to 27% real-world success without wrist cameras, while EyeRobot 2.0 improved on passive stereo by 40% in real tests and 20% in simulation. It also more than doubled ego-plus-wrist policies when grasped objects occluded the wrist cameras, 48% versus 22%. ArXiv · AI/CL/LG's note
EyeRobot 2.0 swivels two eye viewpoints toward a 3D fixation point, then spends more visual computation near the image centers. The system learns gaze control and fixation selection from real-world data, paired with a behavior-cloned gripper policy. In the reported trials, passive stereo fell to 27% real-world success without wrist cameras, while EyeRobot 2.0 improved on passive stereo by 40% in real tests and 20% in simulation. It also more than doubled ego-plus-wrist policies when grasped objects occluded the wrist cameras, 48% versus 22%. ArXiv · AI/CL/LG's note
score 5