Differentiable Simulation for Physical System Identification

Simulating frictional contacts remains a challenging research topic in robotics. Recently, differentiable physics emerged and has proven to be key element model-based Reinforcement Learning (RL) optimal control fields. However, most of the current formulations deploy coarse approximations underlying physical principles. Indeed, classic simulators loose precision by casting Nonlinear Complementarity Problem (NCP) contact into Linear (LCP) simplify computations. Moreover, such methods non-smooth operations cannot automatically differentiated. In this letter, we propose (i) an extension staggered projections algorithm for more accurate solutions problem with friction. Based on formulation, introduce (ii) simulator efficient way compute analytical derivatives involved optimization problems. Finally, (iii) validate proposed framework set experiments present possible application our simulator. particular, using approach demonstrate estimation friction coefficients object masses both synthetic real experiments.