Model-Based Control

Writing out a mathematical model of the robot and its environment first, then deriving or optimizing control commands from it.

A broad category of methods that design controllers around an explicit mathematical model, covering both kinematics and dynamics — for example, the manipulator equation M(q)q̈ + c(q, q̇) + g(q) = τ, where M is the mass matrix, c the Coriolis and centrifugal terms, g the gravity term, and τ the joint torque. Common techniques include gravity compensation, computed torque control (which uses the model to cancel nonlinearities), operational space control, model predictive control (MPC, which optimizes a short future window using the model every cycle), and quadratic-programming-based whole-body control. The advantages are interpretability, no need for training data, and the ability to analyze stability formally; the drawbacks are that the model must be accurate — contact, friction, and deformable objects are hard to model — and complex tasks demand extensive manual design. It sits opposite learning-based control, though the two are now often combined. Note it is distinct from model-based reinforcement learning, where the ‘model’ is itself learned from data.

ExampleMIT Cheetah 3 simplified body dynamics into a convex MPC that solves for ground reaction forces at each foot at 20–30 Hz in under 1 millisecond per solve; the same set of parameters produced trotting, galloping, bounding, and other gaits, reaching forward speeds up to 3 m/s.

Also called

Model-Driven Control

Related

Learning-Based Control · Computed Torque Control · Model Predictive Control · Whole-Body Control · System Identification · Dynamics

Sources

Modern Robotics 11.4: Motion Control with Torque or Force Inputs (Part 3 of 3)（计算力矩控制） (Chinese)
Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control (IROS 2018, MIT DSpace)