Model predictive control

What it means

Model predictive control, or MPC, uses a model to choose actions by solving an optimization problem over a future time window. The problem expresses an objective and constraints on the system's evolution. The controller applies the next action, then repeats the calculation from the new state. Rawlings, Mayne and Diehl develop this approach in their MPC textbook. [1]

The moving window is called a receding horizon. MIT's trajectory-optimization chapter explains that each new calculation considers another finite sequence of future steps. The controller can therefore revise its decisions using feedback instead of committing to an entire previously planned motion. [2]

How it applies to humanoids

Humanoid balance can involve future foot placement, support forces and body motion. A predictive controller can coordinate selected variables ahead of time while accounting for constraints. The push-recovery paper by Shafiee-Ashtiani and colleagues studies an MPC scheme that adjusts balance quantities when stepping is limited. Its evaluation uses an abstract model of SURENA III, rather than establishing a general physical-robot result. [3]

MPC can sit above a lower-level whole-body tracker. Zhang and colleagues use an MPC planner to generate motion and contact-force targets, then a hierarchical optimization layer to track them. Predicting selected body variables is a different job from producing every motor command. Read each layer's model and outputs before describing a system as full-body predictive control. [4]

Terminology and implementation traps

Prediction does not require a learned world model or a language model. MPC's model can be a conventional mathematical description of the plant. The textbook discusses model-based control and optimization; a manufacturer using the word predictive has not thereby identified an AI architecture. [1]

Optimization is not automatically a guarantee of success. MIT explains recursive feasibility: a feasible solution now must not lead to a later problem with no feasible solution. Stability, model uncertainty and solver performance require their own assumptions and analysis. [2]

Horizon length, model detail and update rate should be read together. More elaborate prediction changes the optimization problem and its computational demands. A reported control frequency without the problem definition is insufficient for comparing two implementations. [1, 2]

How to read it on humanoidversus.com

The ai_stack field can retain a named controller if the cited record identifies it, but it is not an MPC performance score. compute describes disclosed hardware, not whether a specific optimization meets its deadlines. This entry has no live table.

No robot links are inferred from walking videos or broad autonomy claims. For a cited MPC result, check whether testing was simulated or physical, what disturbances were used and which body variables were controlled. In particular, a paper using an abstract robot model should not be presented as verified customer capability. [3]

Further reading

The Rawlings, Mayne and Diehl textbook covers MPC theory and computation. MIT explains receding-horizon trajectory optimization and feasibility. The push-recovery paper gives an abstract-model balance example, while Zhang and colleagues describe an MPC planner paired with a whole-body tracking layer. [1–4]