The ability to invert a simulator in order to perform system identification or control is an important tool in the development of any robot that interacts with the world. In the past, roboticists have often taken model-free approaches to problems such as system identification; while sometimes robust and efficient, such methods struggle with reasoning efficiently about the high sensitivity of contact. Further, although gradients have long been used for tasks such as optimal control (e.g. LQR and trajectory optimization), such approaches are often developed in isolation, without considering direct interplay or integration with complex learned models or algorithms.

More recently, there has been growing interest in designing differentiable simulators -- those that can compute ground-truth, analytical gradients of any quantity with respect to any other. These gradients allow direct application of modern numerical first-order (gradient descent) and second-order (Newton or quasi-Newton) optimizers, and thus allow one to solve complex robotics problems more efficiently. For example, differentiable simulation enables direct optimization of closed-loop feedback policies for robots navigating complex and contact-rich environments. As another canonical example, differentiable simulation has been used to match trajectories of corresponding physical and virtual robot models by efficiently searching for parameters that match the model to nature. These capabilities have resulted in breakthrough research results on both simulated and physical hardware.

The development of a differentiable simulator that can be leveraged by algorithms which consume its computed gradients to effectually solve difficult robotics problems is nontrivial. In order to be useful, a differentiable simulator must be fast in both evaluation and differentiation, generalizable across many different problem configurations, resilient to discrete-time events (such as contacts), and have non-vanishing and non-exploding gradients. Further, although the interplay between differentiable simulator and numerical optimization is obvious, direct optimization has its limits, being susceptible to suboptimal local minima. New algorithms which leverage gradients in non-obvious ways are needed to solve robotics' biggest problems in system identification, learning and control, and overcoming the sim-to-real gap.

This workshop seeks to look backward at the biggest successes over the last several years in differentiable simulation, as well as look forward to the biggest hurdles.