Experience says that the models of frictional contact that we use in robotics are noisy. Robots benefit from models that are computationally efficient, accurate, and thoroughly evaluated. Our work focuses evaluating and reinforcing contact models for frictional and impulsive forces. We want to understand when we can trust contact models derived from constitutive laws like Coulomb friction or principle of Maximal Dissipation. We are also interested in using data to reinforce these models and use them in perception, planning and control.
Contact Datasets. We have developed techniques to automate and instrument experiments involving frictional dynamics in a robot arena with precise robot manipulators and carefully calibrated motion tracking and force sensing. This has lead to open datasets of controlled experiments to learn models of frictional sliding and impact.
Data-reinforced models of frictional dynamics.We have shown that experimental data can be used to learn or reinforce models that outperform constitutive laws of friction and restitution, and to capture their inherent variability. We have explored this approach on the dynamics of planar pushing and planar rigid impacts.
Stochastic simulation and planning with data-reinforced models. We are exploring a dynamic filtering scheme GP-SUM, that exploits the algebraic structure of Gaussian Processes to efficiently propagate non-Gaussian beliefs, to simulate state distributions through data-reinforced contact models.