How a four-legged robot stretched one battery across a marathon

A four-legged robot has travelled 42.195 kilometres on one battery—far enough to make legged machines more plausible for long inspections and other journeys where wheels cannot easily go. The interesting part is not that RAIBO2 crossed a marathon finish line. It is how many small energy leaks had to be closed before it could get there.

The KAIST robot finished the Sangju Dried-Persimmon Marathon in South Korea in 4 hours, 19 minutes and 52 seconds. The run happened in November 2024; the peer-reviewed account was published in Nature on September 23, 2026. It used about 1,280 watt-hours during the race—roughly 30.3 watt-hours for every kilometre travelled.

That last number is reproducible with ordinary arithmetic:

1,280 Wh ÷ 42.195 km = 30.3 Wh/km

Dividing the same energy by the finishing time gives an average electrical draw of about 296 watts. These are averages, not a claim that the robot consumed power steadily: hills, speed changes and stops alter the moment-by-moment load.

Where a walking robot loses its charge

Wheels can roll while supporting weight with relatively little continual effort. A quadruped must hold itself up through powered joints, repeatedly accelerate and slow its legs, and survive a small collision every time a foot meets the ground. A slipping foot squanders still more motion.

RAIBO2’s designers treated that collection of losses as one connected problem. They removed material from the legs while retaining rigidity, reducing the mass that had to swing thousands of times. Lower-resistance motor-driver circuitry cut electrical energy dissipated as heat. The saved weight also made room for a larger battery rather than merely producing a lighter robot.

The gait mattered too. A control policy trained with reinforcement learning—a method that rewards useful behaviour during repeated simulations—was encouraged to place the feet with less speed at contact, limit slipping and avoid wasteful electrical loads. Nature’s supplementary description reports treadmill comparisons of policies with and without a collision-related reward, while KAIST’s contemporaneous account describes two approximately 50-metre climbs on the marathon course.

On descents, the motors could act partly as generators, returning some energy to the battery. That is regenerative braking in four-legged form: gravity drives the joints, and the electrical system recovers part of the motion instead of turning all of it into heat.

The researchers have also released marathon logs and comparison data, including battery voltage, current, joint and temperature records. An integration of all valid voltage-current intervals in the supplied logs gives roughly 1,308 watt-hours across about 4.93 hours of logging. That is broadly consistent with the paper’s approximately 1,280-watt-hour race figure, but it is not a replacement for the researchers’ race-window calculation because the archive spans more than the official finishing time.

An efficient runner, not an autonomous one

The paper reports a dimensionless “cost of transport” of 0.25, compared with a human benchmark of 0.37. This measure divides energy by weight and distance, allowing machines and animals of different sizes to be compared. Using the reported 45-kilogram mass gives the same rounded result:

4.608 million joules ÷ (45 kg × 9.81 m/s² × 42,195 m) ≈ 0.247

That comparison does not establish that RAIBO2 used less total energy than a person. The result depends on normalization and on what energy is counted: battery electricity for the robot and metabolic expenditure for a human are not interchangeable accounting systems. The defensible conclusion is narrower and still impressive—the robot moved its own weight over the course with an unusually low measured energy cost for a quadruped. Nature reports more than triple the per-charge range of existing quadrupeds.

Nor was this a demonstration of independent navigation. As Scientific American reported, citing outside robotics researchers, RAIBO2 carried a camera but did not interpret the road itself; a person steered it remotely. Sensors, perception software and onboard navigation would consume power currently available for movement.

The marathon therefore answers one carefully bounded question: can an electrically powered quadruped keep moving for 42.195 kilometres without swapping its battery? RAIBO2 did. Seeing where its energy went—and which losses engineers managed to shave away—is more useful than pretending the finish line proved everything else.