Chaos Physics Engine: Rigid Body Simulation and Deterministic Sub-Stepping: Fundamental Mechanics and Deep Mathematical Modeling

The core of rigid body simulation in Chaos really comes down to how you handle the time integration and contact solving, and most people underweight the integration choice. Chaos uses a fixed-timestep approach with a semi-implicit (symplectic) Euler integrator on the rigid body solver, which is deliberate—explicit Euler drifts energy and blows up on resting stacks, while semi-implicit Euler back-velocity the velocity update before applying position, keeping kinetic energy bounded. For the solver itself, Chaos is impulse-based with a sequential impulse / projected Gauss-Seidel iteration scheme rather than a full constraint matrix solve, which is why you can crank up iteration counts for stability on heavy chains or tall towers.

The key detail everyone misses is that the constraint solver needs warm starting from the previous frame’s impulses—that’s?

Great topic. On the rigid body side, Chaos leans on an impulse-based sequential constraint solver rather than a full LCP formulation—each contact and joint gets resolved iteratively over a fixed solve budget, which is why you see solver iterations and velocity update phases in the r.Physics config. The critical thing people underestimate is the interplay between the broad phase (BVH rebuild cadence) and the narrow phase contact generation: if your broad phase is using discrete swept checks instead of continuous collision detection (CCD), you’ll get tunneling at high angular velocities regardless of how many solver iterations you throw at it. For deterministic builds, make sure you’re pinning your sub-step size as a fixed delta rather than tying it to frame time—Chaos has a bDeterministic flag that forces a fixed accumulator, and if you’re feeding it variable dt you’re going to get frame-dependent divergence that will wreck replay