Robot Learning with Visual Predicted Force
A compliant gripper’s visible deformation is used as the force signal, replacing force or tactile sensors at deployment.
The paper trains a visual force estimator on calibration data, then uses those estimates to add force labels to robot demonstrations. A second policy proposes both actions and the forces they are expected to produce. At test time, the robot chooses the candidate action whose predicted force is closest to the demonstration target. The authors evaluate it on berry picking, empty-can grasping, in-hand reorientation, and plug insertion. ArXiv · AI/CL/LG's note
The paper trains a visual force estimator on calibration data, then uses those estimates to add force labels to robot demonstrations. A second policy proposes both actions and the forces they are expected to produce. At test time, the robot chooses the candidate action whose predicted force is closest to the demonstration target. The authors evaluate it on berry picking, empty-can grasping, in-hand reorientation, and plug insertion. ArXiv · AI/CL/LG's note
score 4