Karen Liu of School of Interactive Computing was awarded the Google Faculty Research Award for 2017. Her project is entitled “Closing the “Reality Gap”: A Machine Learning Approach to Contact Modeling” and aims to accurately compute contact states (sticking, sliding, or breaking) and contact forces such that the simulated results match the real world phenomena. She describes the project as
Accurate physics simulation has become an essential component in motor skill learning for robotics applications. Many machine learning approaches can only be conducted in simulation due to costly and potentially risky data acquisition in real world. However, motor skills learned in simulation often perform poorly on physical hardware due to inaccurate parameters, idealized dynamic and contact models, or other unmodelled factors. This proposal aims to accurately compute contact states (sticking, sliding, or breaking) and contact forces such that the simulated results match the real world phenomena. Our approach constructs a data-driven model that utilizes real-world observations to improve the accuracy of simulation. The key insight is that the contact problem can be broken down to two steps: predicting the next state of each contact point and calculating contact forces based on the prediction and current dynamic state. We solve the first step by learning a classifier from real world data. By doing so, the second step is also simplified to solving a linear system. We will evaluate our approach using both simulated and real world data. As a proof-of-concept demonstration, we will show that a humanoid can perform tasks involving whole-body dynamic balance in real world using the control policy trained by our improved simulator.