Robot Marathon: DARPA Races to Humanoids and ABB

Making a robot run 21 or 42 kilometres on real roads tests most of what robot builders find hard: balance, motors that run for hours without overheating, batteries, and software that copes with uneven ground and crowds. That is why the robot marathon has gone from a curiosity to a serious benchmark. This post looks at how robot races grew out of driverless car challenges and early running humanoids, what happened at Beijing's humanoid half marathon and this year's robot games, and how a four-legged robot finished a full marathon on one charge. It ends with where ABB robots fit and what running robots could mean for homes and factories.

A Unitree G1 humanoid robot at a 2024 expo. Photo: Sayanesy / Wikimedia Commons (CC0)

How robots learned to race

The first robot races to grab public attention were for cars. DARPA's Grand Challenge in 2004 sent driverless vehicles across the Mojave Desert, and none of them reached the finish. A year later Stanford's Stanley, a modified Volkswagen Touareg covered in sensors, won the 2005 race. Every robot race since has repeated what the desert showed: a machine that works in a lab often fails over a long course, because dust, heat, battery drain and small errors pile up hour after hour.

Stanley, the driverless car that won the 2005 DARPA Grand Challenge, on its way to the Smithsonian. Photo: National Museum of American History via Wikimedia Commons (public domain)

Legged robots moved more slowly. RoboCup, first held in Nagoya in 1997, gave university teams a reason every year to build small humanoids that could walk, turn and get up after a fall. Most of that effort went into robot football, but the walking and balance control it pushed forward later fed into running. Kid-size humanoids at RoboCup still tip over often, and anyone who has watched a match can see how hard it is to keep a two-legged machine upright at speed.

Kid-size humanoid robots on the pitch at RoboCup 2016 in Leipzig. Photo: ubahnverleih / Wikimedia Commons (CC0)

Honda's ASIMO took the first steps toward a running humanoid. Introduced in 2000, it could run at 3 km/h by 2004, 6 km/h in a straight line by 2005 and 9 km/h by 2011. That is a slow jog for a person, yet it showed that a full-size humanoid could leave the ground and land again without falling. Honda stopped developing ASIMO in 2018 and retired it in 2022, a few years before Chinese makers began building faster and cheaper humanoids in volume.

Honda's ASIMO on stage at Miraikan in Tokyo. Photo: Davide Mauro / Wikimedia Commons (CC BY-SA 4.0)

Robot marathons and humanoid games today

On 19 April 2025 Beijing's E-Town district held the first half marathon with humanoid robots in the field: 21 robots ran alongside about 12,000 people on a 21 km course. Tiangong Ultra from the Beijing Humanoid Robot Innovation Center, 180 cm tall and 52 kg, won in 2 hours 40 minutes and 42 seconds. Only a handful of robots finished within the four-hour limit, and iRunFar reported that the winner needed three battery changes along the way. One robot fell at the start, and another crashed into a barrier. The fastest human, a runner from Ethiopia, finished in 1 hour 2 minutes and 36 seconds, so the people were still well ahead.

The headquarters of humanoid and robot dog maker Unitree in Hangzhou, April 2026. Photo: Windmemories / Wikimedia Commons (CC BY-SA 4.0)

The second World Humanoid Robot Games ran in Beijing from 22 to 26 August 2026 at the National Speed Skating Oval, with 51 events. Reuters reported that Tiangong Ultra ran 9.39 seconds in a 100 metre heat, faster than Usain Bolt's record. On 23 September Nature published KAIST's RAIBO2, a four-legged robot that finished the Sangju Dried Persimmon Marathon in 4 hours 19 minutes 52 seconds on a single 2,016 Wh battery. An earlier version had run flat at about 37 km, and the team, led by KAIST professor Hwangbo Je-min, estimates RAIBO2 could have gone another 25 km.

A Unitree four-legged robot meets visitors in a Bucharest park in 2023. Photo: Sgt. Amber Edwards / US Army via Wikimedia Commons (public domain)

ABB does not build humanoids or robot dogs, and none of these races used ABB machines. ABB robots could fit in the factories behind them. Running robots need thousands of small actuators, gearboxes and battery packs assembled to tight tolerances, and a dual-arm YuMi IRB 14000 or a GoFa CRB 15000 cobot could plausibly assemble and test such parts beside people. An industrial arm could also run endurance tests on legs and joints. Cells like these are normally planned in RobotStudio before anything is bolted to the floor.

ABB's dual arm YuMi assembling small parts beside a lab worker. Photo: ABB

What robot marathons mean for homes and factories

A race is a public reliability test. A robot that covers 42 km without a battery swap has motors, cooling and software good enough for a long working day, which is what most builders are actually after. KAIST's spin-off Raion Robotics is now working on mass production of quadrupeds, and Chinese humanoid makers bring their machines to races and games where buyers and investors can watch them. For homes, the hoped-for result is a robot that can walk around a house or garden all day without stopping to charge every hour.

A Unitree G1 humanoid, a FOX robot and a Go2 robot dog on stage at Japan Mobility Show 2025. Photo: RuinDig/Yuki Uchida / Wikimedia Commons (CC BY 4.0)

Factories care about the same numbers. A humanoid that walks between workstations has to last at least a shift. KAIST found its earlier robot used more energy than expected because it kept changing speed to adjust to other runners, and steadier motion saves energy in a factory aisle too. Fixed industrial robots have worked on energy use for years, as our post on robot energy efficiency and battery life explains. For many jobs a wheeled base or a fixed arm will still beat a running humanoid on cost and uptime.

Lead-through teaching: moving an ABB GoFa arm by hand. Photo: ABB

Plenty is still unsolved. RAIBO2 had some help from remote human controllers, and most robots in the Beijing half marathon did not finish in time. Falls become a safety problem once robots share pavements, gyms or homes with people. Running humanoids are expensive, and outside a few research groups and start-ups very few people know how to maintain them. Each new race shows these weak points in public, which at least tells builders what to fix next.

An engineer programs a robot cell in RobotStudio next to an ABB industrial robot. Photo: ABB

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com