> QuickAnswer: KAIST's RAIBO2 quadruped robot completed a full 42.195 km marathon on a single battery charge in 4 hours, 19 minutes and 52 seconds. Published in Nature, the result beats the average human marathon finisher. The key breakthrough is not top speed but a reinforcement-learning locomotion algorithm optimised for sustained energy efficiency. RAIBO2 is a research prototype, not a commercial product.
The Run: 42.195 km, One Charge, 4h 19m
On a standard athletics track in South Korea, a four-legged robot completed a full marathon without stopping, without recharging, and without a support crew following alongside with spare batteries. The robot was RAIBO2, built at KAIST, the Korea Advanced Institute of Science and Technology, based in Daejeon, South Korea. The final time was 4 hours, 19 minutes and 52 seconds.
That figure deserves a moment of context. The average human finisher in a mass-participation marathon event takes approximately 4 hours and 30 minutes. RAIBO2 ran a faster marathon than most people who enter one.
The run covered the full official marathon distance of 42.195 kilometres. It was conducted on a track rather than open road, which provides a controlled, repeatable surface and eliminates variables such as gradient and traffic. The results were submitted for peer review and published in Nature, confirming that this was not a carefully edited promotional clip from a company marketing department. It was a scientific result, scrutinised and verified.
Previous robot endurance records over long distances were significantly shorter. This result moves the goalposts considerably, and the robotics community has taken notice.
What Makes RAIBO2 Different
RAIBO2 is a quadruped robot, meaning it moves on four legs rather than wheels or tracks. Quadrupeds can handle terrain that defeats wheeled platforms: steps, rubble, uneven ground, and loose surfaces. But quadruped locomotion has historically been energy-intensive. Moving four legs in coordinated gaits, maintaining balance, and adapting to surface variation all demand significant computation and power.
The headline figure from the KAIST research is not the robot's top speed. RAIBO2 is not trying to sprint. The story here is energy efficiency sustained over a 42-kilometre distance, and the mechanism behind it is a custom reinforcement-learning-based controller.
Reinforcement learning allows a robot to discover locomotion strategies through simulated trial and error rather than following rigid, hand-coded movement patterns. KAIST's team optimised their controller specifically for long-distance running, targeting energy consumption per metre rather than maximising velocity. The result is a gait that keeps power draw low and consistent across the entire run, rather than front-loading effort and degrading performance as the battery depletes.
This is analogous to how elite human marathon runners train to maintain a steady pace and efficient stride economy rather than going out hard and fading. The comparison is not merely poetic. The underlying principle, sustaining output over distance by minimising wasted effort, applies to biological and mechanical locomotion alike.
If you are new to the quadruped robot category, the used robot dog buyers guide covers what these machines are, how they work, and what separates research platforms from commercial products available today.
The Nature Publication
The choice of journal matters. Nature is one of the most selective and widely respected peer-reviewed scientific publications in the world. A result published there has passed scrutiny from independent expert reviewers who have no commercial stake in the outcome.
This distinguishes the RAIBO2 marathon from many robot demonstrations that circulate as video content. There is no shortage of impressive-looking footage of robots doing remarkable things under carefully controlled conditions, with footage selected from the best attempts. A Nature publication requires full methodology, reproducible results, and honest reporting of limitations. The marathon record stands on that basis.
The paper reached mainstream science coverage in early October 2026, bringing the result to a broader audience beyond robotics research circles. That wider attention is warranted. The technical contribution, an energy-efficient locomotion algorithm that enables a quadruped to cover marathon distance on one charge, has implications well beyond academic interest.
How Far Can Commercial Quadrupeds Go?
To understand what RAIBO2 achieved, it helps to look at where commercial quadruped robots currently stand on endurance.
Boston Dynamics Spot is the most recognised commercial quadruped on the market. Under normal operating conditions, Spot's battery provides approximately 90 minutes of runtime. That is sufficient for many inspection and data-collection tasks, but it represents roughly 20 per cent of the time RAIBO2 spent on the track. Boston Dynamics has built Spot into a highly capable platform with a strong software ecosystem and a wide range of payloads, but long-distance endurance was never the design brief.
The Unitree Go2 Pro shares the general quadruped form factor with RAIBO2 and sits at a more accessible price point. Runtime figures vary by payload and terrain, but the Go2 Pro operates for roughly two to four hours under typical conditions. That is meaningfully better than Spot's 90 minutes, but still well short of a four-hour marathon. The Go2 Pro is designed as a versatile commercial platform rather than a long-range endurance machine.
For a direct comparison of how these commercial machines hold their value over time, the resale value analysis for Unitree Go2 vs Boston Dynamics Spot provides useful context on the broader market positioning of each.
Neither Spot nor Go2 Pro was designed for marathon distances, so citing their battery limits as a shortcoming misses the point. What the RAIBO2 result does is reveal that the gap between current commercial endurance and research-level endurance is substantial, which matters when asking what these machines could eventually do in the field.
Why Endurance Matters More Than Speed
The instinct when reading about a robot running a marathon is to focus on the time. But the more important number in the KAIST paper is the distance: 42.195 km on one charge.
Speed is relatively easy to optimise in isolation. A robot can run fast for a short burst. What has historically been difficult is maintaining any useful level of performance across extended distances without access to a power source. Quadrupeds have an inherent disadvantage compared to wheeled robots on flat surfaces because leg-based locomotion requires continuous stabilisation work that wheels do not. Every step demands computation, actuation and balance correction.
KAIST's reinforcement-learning controller addresses this by finding movement patterns that minimise the energy cost of each step rather than maximising stride speed. The result is a robot that does not tire in the conventional sense but also does not waste power. That efficiency, compounded over 42 kilometres, is what makes the achievement significant.
For a look at which quadrupeds combine endurance and capability in the current commercial market, the best robot dogs of 2026 roundup covers the field across different use cases and budgets.
Implications for Real-World Robot Deployment
The practical question is where a 42-kilometre range actually matters.
Disaster response is one obvious context. After an earthquake, flood, or industrial accident, responders need information from areas that may be inaccessible to humans or wheeled vehicles. A quadruped that can cover several kilometres of rubble, return, and still have battery remaining becomes far more operationally useful than one that needs retrieving after 90 minutes.
Agricultural surveillance offers another application. Large farms, forestry operations, and water catchment areas require regular inspection across distances that make manual patrol time-consuming and aerial drones weather-dependent. A ground robot that can walk a full inspection circuit without returning to base for recharging changes the economics of that task.
Long-range perimeter inspection, infrastructure monitoring along pipelines or rail corridors, and search and rescue in wilderness terrain all share the same core requirement: a robot that can cover serious distance under its own power. RAIBO2's result shows that quadruped robots are approaching the range threshold where these applications become practical.
The key caveat is that RAIBO2 ran on a flat, controlled track. Real-world terrain is harder. Energy costs will be higher on grass, gravel, or uneven surfaces. The 42-kilometre figure would reduce in demanding conditions. Even so, the underlying algorithm represents a platform that can be tuned and improved, and a research result on a flat track typically understates what is achievable after further development.
When Will This Reach Commercial Robots?
RAIBO2 is a research prototype developed at KAIST. It is not available for purchase, and there is no announced commercialisation roadmap. The gap between a university research platform and a field-deployable commercial product is typically measured in years and involves significant engineering work around reliability, repairability, software tooling, and safety.
The more immediate impact of the RAIBO2 result is likely to be on the research priorities of commercial robotics companies. When a peer-reviewed paper demonstrates that a quadruped can run a marathon on one charge, it becomes harder to argue that endurance is an unsolvable problem. The reinforcement-learning approach KAIST used is not proprietary in the way a hardware design might be. The methodology is published. Other teams can build on it.
Boston Dynamics, Unitree, ANYbotics, and others working in the quadruped space now have a public benchmark to work towards. Whether the RAIBO2 algorithm translates directly to commercial hardware or inspires parallel approaches, the research will have an influence. How quickly that influence produces field-capable commercial robots depends on factors beyond the algorithm alone: actuator efficiency, battery chemistry, thermal management, and the many unglamorous engineering challenges that separate a laboratory demonstration from a product.
For now, RAIBO2's marathon is a marker. A quadruped robot has covered 42 kilometres on one battery charge, faster than most human runners. That was not possible before. It is now.
Frequently Asked Questions
How fast did the KAIST RAIBO2 robot run a marathon?
KAIST's RAIBO2 completed the 42.195 km marathon distance in 4 hours, 19 minutes and 52 seconds on a single battery charge. This is faster than the average human marathon finisher, which is approximately 4 hours and 30 minutes in mass-participation events.
What is RAIBO2?
RAIBO2 is a quadruped (four-legged) research robot developed by KAIST, the Korea Advanced Institute of Science and Technology, based in Daejeon, South Korea. It uses a reinforcement-learning-based locomotion controller optimised for energy efficiency over long distances rather than peak speed.
Was the RAIBO2 marathon result peer-reviewed?
Yes. The result was published in Nature, one of the world's most prestigious peer-reviewed scientific journals. The methodology and results were independently reviewed before publication.
How does RAIBO2 compare to Boston Dynamics Spot and Unitree Go2?
Boston Dynamics Spot has a rated battery life of approximately 90 minutes under normal use. The Unitree Go2 Pro runs for around 2 to 4 hours depending on load and terrain. Neither is designed for marathon distances. RAIBO2's 4-hour 19-minute run on a single charge represents a significant advance in quadruped endurance, though RAIBO2 is a research prototype not available for purchase.
What are the practical applications of a marathon-capable robot?
Long-range quadruped capability opens applications in disaster response, long-distance perimeter inspection, agricultural surveillance, and search and rescue in terrain where wheeled robots cannot operate. A robot that can cover 42 km without recharging could replace multiple short-range units or reduce reliance on aerial drones in some environments.
Is RAIBO2 available to buy?
No. RAIBO2 is a research platform developed at KAIST. It is not commercially available, and no commercial product based on it has been announced. The research findings are published and available for other teams to build upon.