Hierarchical Control

Splits control into layers: a slow, high-level layer sets goals, and a fast, low-level layer tracks them and drives the motors.

Hierarchical control is a classic approach in control engineering: break a complex control problem into several layers, where the upper layers see farther ahead and update more slowly, and the lower layers see closer and run faster. A higher-level controller outputs reference quantities on a longer cycle — a target velocity, footholds, contact forces, or target joint angles — and a lower-level controller tracks these references at a much higher frequency, handling disturbances and hardware details, generally without the upper layer intervening in how it's executed. The benefit is that each layer can use a simple model and solve its own problem on its own timescale, and layers can be swapped out or debugged independently. It's closely related to 'hierarchical architecture' and 'brain-cerebellum,' which more often describe how task planning and action generation are divided at the model level, while hierarchical control emphasizes the frequencies and interfaces between control loops at each level.

ExampleMIT's Mini Cheetah splits control into two layers: an MPC optimizes foot contact forces over a longer time window using a simplified model, and whole-body impulse control (WBIC) converts those forces into joint commands. Reinforcement-learning locomotion works similarly: a policy outputs target joint angles at 50 Hz, and a lower-level PD controller converts them into torques at a much higher frequency.

Also called

High-Level Controller, Low-Level Controller, Hierarchical Control System

Related

Hierarchical Architecture · Robot Cerebellum · Model Predictive Control · Whole-Body Control · Proportional-Derivative Control · Cascade Control

Sources

Wikipedia: Hierarchical control system
Highly Dynamic Quadruped Locomotion via Whole-Body Impulse Control and Model Predictive Control (arXiv:1909.06586)
legged_gym: legged_robot.py（_compute_torques，PD 控制） (Chinese)