ExTRAORDiNary · Accelerating Differentiable Robot Dynamics Simulation for Advanced Control

Differentiable robot dynamics simulation is a crucial enabler of advanced robot control. It is at the heart of both model predictive control (MPC) and learning-based approaches (e.g. reinforcement learning [RL]), which are among the most successful and actively researched robot control algorithms. Increased usage of the computationally demanding MPC/RL controllers has led to a growing need for efficient dynamics simulators. However, existing simulators internally use inefficient high-complexity (worst-case cubic) constrained dynamics algorithms (CDA) and are often inefficiently implemented leading to a slowdown of several factors compared to a fast simulator like Pinocchio.Addressing these concerns, we will accelerate the differentiable simulation through three complementary strategies. We will 1) leverage low-complexity CDAs, 2) use Pinocchio's proven efficient software design patterns and explore further acceleration via code generation computations, and 3) derive efficient algorithms for differentiating through contact simulation.Furthermore, our simulator will solve the nonlinear complementarity problem of frictional contact without making physics-compromising relaxations like existing simulators and will be publicly available as part of the widely used open-source Pinocchio library. By adding key enhancements to Pinocchio, we will make it a viable alternative to the inefficient, but feature-rich software simulators. The visibility, impact and usability of our simulator will be enhanced by addressing some low-hanging fruits in MPC, RL and physics identification applications.This projects contributions will not only pave the way towards fast whole-body controllers and faster and more sustainable RL training (important a time surge of RL research activity), but will also impact adjacent fields like bio-mechanics and computer graphics in the long term