Learning to predict robots’ motion through impact is hard—but how hard? and why? We isolate near-discontinuity, driven by material stiffness, as a key culprit; and we study how 3 distinct phenomena contribute to poor prediction at test time: chaos, under-fitting, and poor generalization.
Paper - Website - Video - Code
Read MoreCommon methods for learning robot dynamics assume motion is continuous, causing unrealistic model predictions for systems undergoing discontinuous impact and stiction behavior. We resolve this conflict by implicitly encoding these discontinuities as inter-body signed distance and contact-frame Jacobians. Our method, ContactNets, can predict realistic impact, non-penetration, and stiction when trained on 60 seconds of real-world data.
Paper - Video - CoRL Talk (5 min) - Code
Read More
