Model Predictive Control (MPC) Explained for Robot Walking

Planning a few steps ahead, recalculating constantly, rather than committing to one fixed plan.

Model predictive control solves a short-horizon optimization problem — typically predicting and optimizing motion a fraction of a second to a couple of seconds into the future — using a physical model of the robot, then executes only the very first step of that plan before immediately re-solving with fresh sensor data.

Why re-solving constantly matters

Real-world conditions never match a plan perfectly — the ground might be slightly uneven, a push might land unexpectedly. By continuously re-optimizing rather than committing to a long fixed plan, MPC naturally adapts to those disturbances without needing a separate correction mechanism bolted on afterward.

The tradeoff

Solving an optimization problem many times per second is computationally expensive, so MPC controllers need either a fast solver, a simplified physical model, or both — usually some combination of the two — to run at the update rates walking control demands.