Gait Planning

Deciding when each of a legged robot's feet touches down and lifts off, and the rhythm of its steps.

Gait planning decides the contact timing of a legged robot's legs: within one gait cycle, when each leg touches down (stance phase) and when it lifts off (swing phase), what fraction of the cycle each leg spends on the ground (duty factor), and the phase offsets and stepping frequency between legs. Different combinations correspond to different gaits: for a quadruped, diagonal pairs together is trotting, same-side pairs together is pacing, front and rear pairs together is bounding, and all four legs lifting and landing together is pronking. In model-based control, a gait scheduler first lays out this contact timetable, MPC then allocates ground-reaction force to the supporting legs accordingly, and footstep planning decides where the swinging legs should land. Reinforcement-learning locomotion controllers commonly use a phase-clock signal or reward terms to guide the gait, or let a gait emerge on its own. It complements footstep planning: one decides when to step, the other decides where.

ExampleMIT Cheetah's open-source code divides one gait cycle into 10 segments: a trot is written as the four legs having phase offsets (0,5,5,0), each supporting for 5 segments — meaning diagonal legs move together, and each leg is on the ground half the time — while a pronk has all four offsets at zero. Walk These Ways (2022) turns these phase offsets into a policy input command, with (0.5,0,0) giving a trot and (0,0.5,0) giving a bound.

Also called

Gait Scheduling, Gait Generation

Related

Gait · Gait Planning · Gait Cycle and Duty Factor · Footstep Planning · Convex MPC · Central Pattern Generator

Sources

MIT Cheetah-Software ConvexMPCLocomotion.cpp（步态偏移与时长定义） (Chinese)
Walk These Ways: Tuning Robot Control for Generalization with Multiplicity of Behavior (arXiv 2212.03238)
Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control (IROS 2018)