End-to-End Differentiable Model of Robot-Terrain Interactions

Abstract

We propose a differentiable model of robot-terrain interactions that delivers the expected robot trajectory given an onboard camera image and the robot control. The model is trained on a real dataset that covers various terrains ranging from vegetation to man-made obstacles. Since robot-endangering interactions are naturally absent in real-world training data, the consequent learning of the model suffers from training/testing distribution mismatch, and the quality of the result strongly depends on generalization of the model. Consequently, we propose a~grey-box, explainable, physics-aware, and end-to-end differentiable model that achieves better generalization through strong geometrical and physical priors. Our model, which functions as an image-conditioned differentiable simulation, can generate millions of trajectories per second and provides interpretable intermediate outputs that enable efficient self-supervision. Our experimental evaluation demonstrates that the model outperforms state-of-the-art methods.