Simulate robots, backpropagate through collisions

What it does

Tiny Differentiable Simulator is a header-only C++ (and CUDA) physics engine with zero dependencies. It simulates rigid-body dynamics—forward and inverse—along with contact resolution via impulse-level LCP or nonlinear spring-dampers, plus actuator models for motors, servos, and series-elastic joints. The whole library is templatized so you can drop in automatic-differentiation scalars like CppAD or Stan Math types, or stick with plain float and double; it even bundles fixed-point integer math if you need deterministic cross-platform results.

The interesting bit

Because every operation is templated on the scalar type, the same codebase that runs your real-time visualization can also emit forward- or reverse-mode physics gradients for thousands of parallel GPU simulations. That turns the simulator itself into a differentiable layer you can train through.

Key highlights

• Header-only, zero dependencies, C++17.
• Supports forward- and reverse-mode AD via CppAD, Stan Math fvar, ceres::Jet, or standard numeric types.
• Includes fixed-point math for deterministic, cross-platform reproducibility.
• Runs thousands of parallel sims on a single RTX 2080 at 50 FPS.
• Built-in URDF parser and two visualizers: native OpenGL 3+ and web-based MeshCat via ZMQ.

Verdict

Grab it if you are building differentiable robotics pipelines or sim-to-real RL and need physics gradients without dragging in a heavy framework. Skip it if you just need a game-engine physics drop-in with artist-friendly tooling.