EyeRobot 2.0 swaps wrist cameras for a gaze-fixating stereo rig

Researchers from UC Berkeley, Stanford, and Amazon FAR put a single swiveling stereo rig in place of the wrist-mounted cameras that most bimanual manipulation setups rely on, and found it holds up where the standard setup falls apart.

In the arXiv preprint, the authors write that EyeRobot 2.0 "physically attends to a 3D fixation point in the scene by swiveling two eye viewpoints to center their gaze on it." The paper covers over 1,000 physical trials across seven real-world tasks and 1,800 simulated trials across six.

The headline number is in the occlusion case. When grasped objects blocked wrist cameras on tasks like boba and pot lid, the active-gaze system succeeded on 48% of trials against 22% for the ego-plus-wrist baseline. On tasks where wrist views were unobstructed, the two were close: 69% to 64%.

Aggregated across 640 real-world trials, EyeRobot 2.0 scored 67% to the wrist baseline's 52% and the passive-stereo baseline's 27%. The authors describe the architecture as "fixation-centric multi-resolution crops, preserving detail at the center and providing larger context at the periphery," trained with a two-stage RL pipeline that "first train[s] a gaze controller which can look at each per-task object, then learn[s] a gaze sequencing policy."

Kush Hari and Justin Kerr are listed as co-first authors. Ken Goldberg and Angjoo Kanazawa share advising credit. The project page lists no per-task success numbers for the full seven-task real suite.