How RAIBO2 learned to make a battery last for a marathon

The people accompanying RAIBO2 shared the work in 10-kilometre segments: controlling the robot, operating a camera and managing its passage through the crowd. The four-legged machine completed the whole marathon with one battery. On 17 November 2024, it covered the Sangju Gotgam course in South Korea in four hours, 19 minutes and 52 seconds.[1]

The engineering explanation arrived in Nature on 23 September 2026. Researchers led by KAIST’s Jemin Hwangbo approached endurance as a problem spread across the robot’s body, power electronics and learned movement. Their result makes a surprisingly small moment worth examining: the instant a foot reaches the ground.[2][3]

Before the foot hits the ground

Every step has an energy bill. A moving foot carries kinetic energy; an abrupt landing can dissipate some of it. Meanwhile, the motors must support the robot’s weight and swing its legs, with electrical resistance turning some of the supplied energy into heat. A movement that looks smooth from a distance can still be expensive in either place.

RAIBO2’s controller was trained in simulation to reduce energy losses while following movement commands. Its rewards included slower foot motion immediately before contact and lower electrical losses in the actuators. At four metres per second on a treadmill, removing the foot-collision reward substantially increased mechanical losses.[4][2]

That experiment asks a more revealing question than whether the robot can stay upright. If an energy term is removed while the hardware remains the same, the resulting gait shows what the learning rule contributes. The team’s current public code preserves separate variants without the actuator Joule-loss reward, foot-collision reward or ground-force smoothing reward.[5]

The physical intuition is familiar to anyone who has tried to set down a heavy object gently. Slowing it before contact changes how much energy the collision must absorb. A robot has to make that adjustment repeatedly while moving, balancing and responding to its commanded speed. The useful skill is economical contact, sustained over thousands of steps.

Hardware work addressed the other side of the bill. The team reduced unnecessary structural material and investigated losses in the motor-driving electronics, including current sensing, circuit resistance and transistor switching. Public supplementary descriptions also document treadmill comparisons of different body heights: posture changes the balance between mechanical and electrical losses.[2]

These decisions interact. A component that wastes less electricity is helpful, but its mass must also be carried. Raising a body changes the leverage at the joints and the mechanics of landing. Optimising an isolated part therefore offers an incomplete answer to the practical question: how much battery energy does the whole machine spend getting somewhere?

The energy budget for 42.195 kilometres

The reported consumption amounts to about 63.5% of nominal battery capacity. That gives the finish an additional meaning: the machine reached it with an energy margin. Converting all remaining nominal capacity into a guaranteed extra distance would require assumptions about usable battery energy and subsequent conditions; the completed marathon is the observed result.[3]

The authors have also released a dataset containing marathon logs and a separate archive of policy-ablation measurements. The marathon download includes robot records with battery voltage and current, alongside a GPS track. Those are the ingredients behind an energy-and-distance account, although reproducing the race total requires aligning the recording intervals with the event.[6]

For a field robot, a lower movement bill can buy several things: a longer route, more time using instruments, or a larger reserve for the return journey. The right choice depends on the job. RAIBO2 supplies a concrete engineering lesson for all of them: battery life can improve when the machine learns to waste less energy at the point where its body meets the world.

Report an error or suggest a correction