1: \begin{abstract}

2: %While simulators are able to model increasingly complex dynamical systems, they only implement a simplified view of reality. In addition, it is often difficult to find the correct simulation parameters that allow the model to reproduce measurements obtained from the real system, leading to the well known sim-to-real gap.

3: 

4: %In this work we tackle the problem of parameter uncertainty where we aim to infer a probability distribution over simulation parameters given real-world measurements from the actual system. The estimation of such parameters not only enables down-stream applications in transferring controllers from simulation to reality, but the uncertainty information can further benefit robust and safety-critical control applications.

5: 

6: %Leveraging a highly parallel differentiable simulator, we develop a Bayesian parameter estimation technique based on the Stein Variational Gradient Descent method. Such particle-based inference algorithm leverages massive parallelism on the GPU to evaluate the likelihood and its gradient over hundreds of parameter guesses simultaneously. For the likelihood evaluation, we propose a multiple-shooting formulation that significantly improves convergence compared to the commonly used single-shooting objective.

7: 

8: %Our particle-based algorithm accurately infers nonparametric distributions over simulation parameters. We demonstrate on several real-world experiments the scalability of our approach and evaluate the correctness of the found parameter distribution against several other particle-based inference algorithms.

9: %Our gradient-based estimation pipeline is particularly more efficient on high-dimensional parameter inference problems, as we demonstrate on a planar pushing scenario where a Bayesian neural network augments the simulator to learn a detailed model of contact friction.

10: %\end{abstract}

11: