Leveraging Differentiable Simulation with Learned Dynamics Models

Determine how to best leverage differentiable simulators and augment them with learning-based dynamics models to improve optimization and sim-to-real transfer in robotic control tasks.

Background

Differentiable simulators provide analytical or automatically computed gradients that can accelerate and stabilize optimization compared to finite-difference methods, enabling end-to-end training for control and reinforcement learning. Multiple frameworks (e.g., Warp, Taichi, JAX, PyTorch compilation) offer practical tooling.

The authors highlight uncertainty about practical integration strategies: how to combine differentiable physics with learned residuals or data-driven components to achieve robust transfer to real systems.