KAIST’s RAIBO2 ran a 42.195 km marathon on one charge; Nature paper explains how

KAIST’s RAIBO2 four-legged robot ran the full 42.195 km marathon distance in 4 hours 19 minutes 52 seconds on a single battery charge, and the paper describing how was published in Nature on September 23, 2026. The run took place at the Sangju Dried Persimmon Marathon on November 17, 2024; the research was led by Professor Hwangbo Jemin of KAIST’s Department of Mechanical Engineering.

Why it matters: range, not agility, is what limits legged robots in inspection, security and disaster work. The paper shows a roughly threefold gain in range came from engineering out losses across the whole machine, which is a method other robot makers can copy.

How did RAIBO2 run a marathon on one battery?

By cutting energy losses everywhere rather than adding battery. The paper describes a system-wide optimisation of hardware and software. The legs use a lightweight, force-transparent structure, with material removed from the hip mounts and calf links where stress analysis showed it was not needed. Needle bearings and crossed roller bearings reduce friction in the joints. The motor driving circuit was redesigned for low resistance, with an optimised MOSFET arrangement, to cut electrical losses.

Control is the second half. The gait is a reinforcement learning policy trained in simulation with proximal policy optimisation and a curriculum across three levels of terrain difficulty. It was shaped to dissipate little energy: the team modelled two loss sources in particular, the energy spent holding the body up through the joints and the kinetic energy lost each time a foot strikes the ground. Seoul Economic Daily reports the controller was tuned to minimise landing impact and slippage. The approach builds on the simulation-trained locomotion methods explained in our sim-to-real primer.

What is new in the RAIBO2 Nature paper?

The new element is treating endurance as a co-design problem rather than a battery problem. Mechanical design, power electronics and the learned gait were optimised together against one measure, energy per distance. The reported cost of transport, a dimensionless measure of energy used per unit weight per distance, is 0.25. The paper compares that with 0.37 for human runners, meaning RAIBO2 used less energy per kilogram of body weight than a person covering the same ground.

The race itself used about 1,280 Wh, and the team estimates a full battery would carry the robot about 65 km. Earlier quadrupeds typically managed around 20 km per charge, according to Interesting Engineering’s summary of the paper. The result was not a first try: a September 2024 attempt ended at around the 37 km mark. The average pace over the full distance works out to about 2.7 m/s, or 9.7 km/h. The Korea Times and Seoul Economic Daily report that this is the first Korean robotics study published in Nature’s main journal.

Will RAIBO2 be sold commercially?

Yes, through Raion Robotics, a spin-off from Hwangbo’s lab. Seoul Economic Daily reports the company sold nine RAIBO2 units last year and aims to mass-produce 50 quadrupeds in 2026. Target uses are disaster response, perimeter security and hazardous industrial inspection. “At sites such as mountainous or disaster areas where it is hard to secure a charging location, range can limit the scope of their use,” the research team said, according to Seoul Economic Daily. Hwangbo told the Korea Times the aim is to “turn the world-class performance achieved in the lab into products.”

Fifty units is small, but it is a product that ships, which is the test our ranking of Korea’s physical AI companies applies. For other robot race records, and what it costs to hire robots for events, see our guide to humanoid robot rental costs.