Skip to main content

Listed

, home
FREN

Model predictive control

In French:
Commande prédictive
Abbreviation:
MPC

Definition

A control method that, at each instant, computes by optimisation the best sequence of commands over a short time horizon, from a model of the system. Only the first command is applied, then the calculation starts again with the new measurements. It assumes the optimisation is fast enough to be repeated continuously during the movement.

Checked on against the references below. · Written with AI · checked against the sources Source: Russ Tedrake, Underactuated Robotics (MIT), chapter “Trajectory Optimization”.

Concrete examples

A GitHub repository offers a convex model predictive controller for the Unitree Go2. The prediction horizon spans one full gait cycle, split into 16 time steps, and a convex quadratic program is solved with OSQP at every predictive control update, about 30 to 50 times per second. The results published in the repository come from MuJoCo simulation.

What it changes for you

Computation cost forces trade-offs: according to Rathod et al., a full dynamics model over a long horizon can demand too much computation for real-time use at high rates, hence approximate models, at the price of accuracy. In that work, the predictive controller tested on the HyQ quadruped re-plans at 25 Hz with a 2-second horizon and uses a single rigid body model to reduce computation.

Additional material checked on .

Related terms

References

Back to the glossary