Extended Data Fig. 2: Additional evaluation on visuomotor control.
From: Controlling diverse robots by inferring Jacobian fields with deep networks
(a, b), Our framework controls a 3D-printed pneumatic hand to grasp on an object and complete a pick-and-place task. (c, d), Our approach controls the Allegro hand to close and open each finger fully. (e, f), We intentionally perturb the scenes to create occlusion and background changes. We find that the Jacobian field is robust against out-of-distribution scenarios and successfully controls the hand to achieve detailed motions.
