Extended Data Fig. 6: Visual Jacobian comparison against analytical simulations.
From: Controlling diverse robots by inferring Jacobian fields with deep networks
a, We evaluate the quality of our Jacobian predictions against the analytically computed Jacobian field for the Allegro hand in Drake47. We find that our model highly matches the analytical Jacobian field, including the low-to-high opacity transition on each finger, which indicates the total change of position upon a command. We highlight that computing the analytical Jacobian relies on the existence of a physics simulator and a URDF file. b, We compare each component of the predicted Jacobian (red arrow), queried at a 3D position, with its analytical counterpart (blue arrow). Each column visualizes a different command channel inside the Jacobian, covering all of the robot’s motors. Our model achieves high-quality Jacobian predictions, learning only from video observations.
