Extended Data Fig. 5: Qualitative comparison on test set between our Jacobian model and the direct neural flow baseline.
From: Controlling diverse robots by inferring Jacobian fields with deep networks
We evaluate our model and the baseline on the testing samples reported in Extended Data Table 3. Consistent with the numerical results, we qualitatively find that our model can predict correct optical flows on the testing dataset. In comparison, the baseline optical flow model fails to explain out-of-distribution robot commands due to the lack of inductive biases on the locality and symmetry of the dynamical system. In the last row, we visualize the components of our Jacobian model. This validates that the Jacobian model can break down the spatial volume into parts sensitive to each robot finger on the testing dataset. The Jacobian coloration scheme is consistent with Fig. 2.
