EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras
Automatically prepared brief
EyeRobot 2.0 uses active gaze to enable precise bimanual manipulation with a single stereo camera, improving performance over traditional methods without wrist cameras.
The research question addresses how to achieve precise bimanual manipulation using only a single stereo camera instead of multiple cameras, including wrist cameras, to reduce complexity and cost.
The proposed approach involves EyeRobot 2.0 actively focusing its gaze on a 3D point in the scene by adjusting its two viewpoints, processing these images with increased focus on central areas, and using reinforcement learning to coordinate gaze and select targets based on task progress.
The authors report evaluating EyeRobot 2.0 against passive stereo and ego + wrist camera policies using both real-world and simulated data. EyeRobot 2.0 outperforms passive stereo and matches or exceeds ego + wrist policies in various scenarios.
About this brief
This article summarizes a public preprint abstract. Its findings have not been independently reproduced for this publication. Consult the paper for methods, evaluation scope and limitations. Automated briefs may contain errors; the original source takes precedence.