Event

PointHand Keypoint Rendering Improves Novice Teleoperation Demonstration Quality

1 reports · 1 sources

Amane Yamaguchi et al. propose the PointHand teleoperation interface, which renders a multi-fingered robotic hand as sparse keypoints at the fingertips and joints (consistent with the keypoints observable by the policy). In a simulated cup-into-box task, they compare three hand rendering methods (realistic hand shape, all-keypoint emphasis, and thumb-and-index-finger-only emphasis) for their effects on novice demonstration quality and the success rate of diffusion policies trained on those demonstrations. A total of 18 novices and 4 experts collected 2,200 trials; among them, novices who saw the thumb-and-index-finger emphasis rendering achieved 441 successful demonstrations out of 600 trials, higher than the 320 successes with realistic hand-shape rendering. The paper reports that after training diffusion policies on these demonstrations (replacing the policy's observed scene point cloud and proprioceptive state with hand keypoints and two object keypoints), this condition went from the weakest among the three novice conditions under the point-cloud policy (46.3%) to the strongest under the keypoint policy (57.3%) within 400 steps, an improvement of +11.0 percentage points across all five random seeds.