Accurate physics simulation has become an essential component for developing robots that physically interact with the world. A particularly important aspect is simulating contacts between the robots and objects in the environment, which can be useful for both planning and machine learning. However, robots that learn in simulation often perform poorly in the real world due to inaccurate parameters, idealized dynamic and contact models, or other unmodeled factors. This project therefore tackles an important challenge in physics simulation: accurate modeling of contacts. The results will significantly improve contact modeling in physics simulation, which offers a safe space to learn difficult and highly risky motor skills such that the robots can operate more efficiently and robustly even in unseen scenarios in the real world. This capability will have potential impact on robotics in healthcare, search-and-rescue, and space exploration.
This proposal introduces a technique that effectively utilizes real-world data to model the complex, poorly understood contact phenomena. Specifically, the new data-driven technique accurately computes contact states (sticking, sliding, or breaking) and contact forces such that the simulated results match the real-world phenomena. Instead of taking the conventional approach of system identification, this proposal leverages empirical evidence and deep learning techniques to enhance the existing contact model, namely, an implicit time-stepping, velocity-based Linear-Complementarity Program (LCP). The key insight is that the contact problem can be broken down into two steps: predicting the next state of each contact point and calculating contact forces based on the prediction and current dynamic state. The first step is solved by learning a classifier from real-world data. By doing so, the second step can be simplified from an LCP to a Linear Program (LP), thus making the calculation of contacts much more efficient.
