Unitree's Humanoid Robot Now Walks Using 16% Less Power

Walking robots are famously power-hungry. Unlike a car rolling on wheels, a two-legged robot has to constantly catch itself from falling, and its motors burn electricity doing that every single step. A team of robotics researchers has now shown a way to make that walking noticeably cheaper — using a trick borrowed from how humans and toys naturally move.

The idea comes from "passive dynamic walking." A simple wind-up toy can walk down a slope using no motor at all, because gravity does the work and the legs swing like pendulums. The researchers copied that condition inside a computer simulation: during early training, they tilted the pull of gravity so the robot felt like it was always walking slightly downhill. That let the robot stumble onto an efficient, gliding gait by itself. Crucially, they then removed the tilt entirely and let the robot keep practicing under normal conditions, so the final skill is the robot's own, not a crutch.

The test subject was Unitree's G1, a 29-joint humanoid about the size of a small adult. Across walking speeds from a slow stroll (0.5 m/s) to a brisk jog (2.0 m/s), the method cut the energy needed to move a given distance by roughly 7% to 15%, without making the robot any slower or sloppier at hitting its target pace. On physical hardware the savings reached 16.3% when paired with a basic walking pattern. Without that pattern, the gain was a modest 4.5% — small enough to be within normal run-to-run variation, so not yet a guaranteed win.

Why care? Battery life is the single biggest brake on useful robots. Every percent of energy saved is extra minutes of work, a smaller battery pack, or a lower electricity bill for whoever owns the fleet. This technique is also essentially free — it's a change to how you train the robot, not extra hardware. The catch: it was demonstrated on one robot model, in largely simulated settings, and humanoids in warehouses and homes are still years from everyday life.