Legged Robots That Balance

Listen — short summary

0:00 / 3:16

The premise of this 1986 book is almost too clean: if you want to build a machine that runs and balances, start with one leg. Not because one leg is simpler in any obvious engineering sense — a pogo stick with a computer strapped to it is genuinely hard to control — but because one leg forces you to confront the real problem. Balance isn't something you solve once and then add locomotion on top. They're the same problem.

Raibert built physical machines, not simulations. The planar hopper, the 3D hopper, the quadruped — hydraulically actuated robots in actual labs — and the key finding from each was the same: you can decompose the problem. Hop height, forward velocity, and body attitude can be controlled mostly independently. This is not obvious from first principles; it falls out of the dynamics when you actually build something and measure it. The "neutral point" for foot placement — the position at which the body neither accelerates nor decelerates — turns out to be straightforward to compute, and once you have it, running is mostly a matter of putting the foot down in the right place at the right time.

The extension to multiple legs is where the book gets philosophically interesting. Raibert introduces the "virtual leg" abstraction: when multiple legs are on the ground simultaneously, treat them as if they were one leg at the center. This is an enormously productive simplification, and it's why the quadruped experiments in the final chapters work as well as they do. The gait patterns — trot, bound, pace — all emerge from this framework with the same basic control laws, just with different leg phasing. You don't need separate control strategies for separate gaits; symmetry does the work.

The weakest section is the chapter on tabular control — using lookup tables rather than analytic controllers to generate running behavior. It reads like an experiment that didn't pan out, included because it was done. Raibert is honest about its limitations, which is worth something, but it doesn't connect cleanly to the rest of the argument.

What the book understands that most robotics research of that era didn't is that stability in locomotion doesn't mean static stability. A bicycle falls over if it stops moving. A running animal is never statically stable — it's balancing dynamically, exploiting momentum and the periodicity of its gait. Raibert's machines are all "dynamically stable," which means they can fall if the control fails, and the control never lets them. That's a different engineering problem than building something that can stand still, and it's closer to how legs actually work in biology.

This book came out four decades ago and the hardware is primitive by current standards — the quadruped is tethered, the control loops are slow, the sensors are crude. But the conceptual work here is where modern legged robotics begins. The decomposition principle, the virtual leg, the foot placement algorithm — these ideas run through every serious legged robot built since. Read it for the thinking, not the specs.

Key takeaways

Balancing a running one-legged machine requires only three decoupled control loops — hopping height, body attitude, and forward speed — not a single coupled controller managing all variables at once.
Foot placement at the 'neutral point,' a position calculated from current body velocity, is the primary mechanism for maintaining balance across a running stride.
Dynamic, active balance during running is no harder to achieve than static balance; Raibert's machines accomplish it with surprisingly simple feedback laws.
The one-leg control strategy scales directly to multi-legged gaits by treating each pair of simultaneously-grounded legs as a single 'virtual leg.'
Running gaits are governed by spatiotemporal symmetry, which explains why physically different animals converge on a small set of similar gait patterns.
The leg must store and return energy like a spring during stance; without passive energy recycling through leg compliance, dynamic running is energetically unsustainable.
Decoupling the locomotion problem — solving balance, speed, and posture independently — is what makes legged robot control tractable, and it mirrors how biological motor systems appear to be organized.