Long-Horizon Prediction and Uncertainty Propagation with Residual Point Contact Learners

The ability to simulate and predict the outcome of contacts is paramount to\nthe successful execution of many robotic tasks. Simulators are powerful tools\nfor the design of robots and their behaviors, yet the discrepancy between their\npredictions and observed data limit their usability. In this paper, we propose\na self-supervised approach to learning residual models for rigid-body\nsimulators that exploits corrections of contact models to refine predictive\nperformance and propagate uncertainty. We empirically evaluate the framework by\npredicting the outcomes of planar dice rolls and compare it's performance to\nstate-of-the-art techniques.\n