Researchers built a robot that swivels its eyes to stare down whatever it's grabbing, so it no longer needs a camera bolted to its wrist.
EyeRobot 2.0 is a framework described in a new arXiv paper that gives a robot two swiveling eye cameras instead of relying on a fixed wrist-mounted one. The system fixes its gaze on a 3D point in the scene, then processes that image with more computational detail at the center, mimicking how human eyes work. A low-level policy aims the cameras at a target object, while a higher-level policy decides what to look at next based on how the task is progressing; both were trained with reinforcement learning on real-world teleoperation data across 7 physical and 6 simulated tasks. Across more than 1,000 real-world and 1,800 simulated trials, EyeRobot 2.0 beat a standard stereo-only setup by 40% in the real world and 20% in simulation, and matched ego-plus-wrist-camera performance when the wrist view was unobstructed, 69% versus 64%.
Wrist cameras are a known failure point in manipulation: when researchers simply removed them from a standard policy, real-world success collapsed from 52% to 27%. EyeRobot 2.0's standout result is that when a grasped object occluded the wrist camera, it still more than doubled that setup's success rate, 48% versus 22%, suggesting active gaze is a sturdier substitute than just bolting on more fixed cameras.
It is a reminder that better robot vision is not always about adding more cameras; sometimes it is about aiming the ones you already have more intelligently.