Extended Data Fig. 3: Qualitative results on robustness against out-of-distribution scenarios.
From: Controlling diverse robots by inferring Jacobian fields with deep networks
a, We perturb the visual scenes by placing objects around the pneumatic hand mounted on the robot arm. Consequently, the input observation is highly out-of-distribution from the training data. We visualize the predictions of depth and Jacobian at both the input and novel viewpoints. We find that our method retains high-quality predictions. (b, c), We place objects around the pneumatic hand to create occlusion. For presentation clarity, we overlay the Jacobian prediction on RGB images to highlight the shapes of the hand with masking.
