Foundations of Motion Prediction (Dynamic Simulation)

Motion prediction in the presence of intermittent contact between objects is a key component of manipulation planning. Both classical approaches and learning-based approaches for manipulation planning and control need motion prediction methods. The goal of this research is to understand the foundational aspects of motion prediction for objects in contact. Two fundamental physical constraints that two solids undergoing relative motion in contact with each other satisfy are: (a) the solid objects do not interpenetrate (b) there is energy dissipation at the contact. Unfortunately, most current dynamic models (especially in discrete-time) and algorithms (or their software implementation) for dynamic simulation violate the physical constraint of non-penetration. This along with other approximations made with the view of making simulation through contact computationally faster leads to many artifacts in dynamic simulation like lack of conservation of energy and inability to model scenarios with non-point contact in a rigorous manner. Thus, it is unclear how much trust one can put into the output of simulations and whether the mismatch between the simulations and reality is due to modeling errors, parameter errors, or the fact that many non-physical approximations are made in the simulation approaches. We study rigorous methods for dynamic modeling and simulation that ensures that the basic physical constraints of non-penetration and energy dissipation at contacts are satisfied. This leads to dynamic simulation algorithms that respects basic physics and allows us to simulate scenarios without introducing non-physical artifacts. Thus, it allows us to predict motion more accurately and also predict motions in situations with non-point contact that are not possible with current algorithms. Please see below for more details, videos, and related publications.