ContactNets: Learning Discontinuous Contact Dynamics with Smooth, Implicit Representations

Common methods for learning robot dynamics assume motion is continuous,\ncausing unrealistic model predictions for systems undergoing discontinuous\nimpact and stiction behavior. In this work, we resolve this conflict with a\nsmooth, implicit encoding of the structure inherent to contact-induced\ndiscontinuities. Our method, ContactNets, learns parameterizations of\ninter-body signed distance and contact-frame Jacobians, a representation that\nis compatible with many simulation, control, and planning environments for\nrobotics. We furthermore circumvent the need to differentiate through stiff or\nnon-smooth dynamics with a novel loss function inspired by the principles of\ncomplementarity and maximum dissipation. Our method can predict realistic\nimpact, non-penetration, and stiction when trained on 60 seconds of real-world\ndata.\n