Robot dog ran a marathon on one charge and lived to publish the paper
Two years later, creators explain how lightweight hardware and reinforcement learning helped it go the distance
In a remarkable achievement, researchers from the Korea Advanced Institute of Science and Technology (KAIST) have published findings detailing how their four-legged robot, RAIBO2, successfully completed a marathon on a single battery charge. The marathon, known as the Sangju Dried Persimmon Marathon, spanned 42.195 kilometers (26.2 miles) and was completed by RAIBO2 in 4 hours, 19 minutes, and 52 seconds.
This marks the first time a quadruped robot has successfully finished a full marathon at an official event, as previously reported by The Register.
The KAIST team, led by associate professor Jemin Hwangbo, managed to significantly enhance RAIBO2's range per battery charge, achieving more than three times the distance compared to recent comparable quadrupeds. While humans currently hold the world record for the fastest marathon time at 1 hour, 59 minutes, and 30 seconds set by Kenya's Sabastian Sawe, quadruped robots like RAIBO2 have unique advantages in navigating uneven terrain, such as stairs, rubble, slopes, and other obstacles that are more challenging for wheeled robots.
The team achieved this remarkable feat by combining lightweight mechanical hardware, low-loss motor-driving circuitry, and an energy-optimized locomotion policy trained through reinforcement learning. RAIBO2 completed the marathon at an average speed of 2.64 meters per second, navigating 286 meters of elevation change and slopes reaching 18.4 degrees.
The robot consumed 1,280 watt-hours from a battery with a capacity of 2,016 watt-hours, resulting in a total cost of transport of 0.25, which is lower than the researchers' benchmark for humans of 0.37.
Despite this groundbreaking efficiency, RAIBO2 is not yet autonomous. The researchers note that the robot still requires human operation for stairs or mountainous terrain, as variations in the energy loss models and constrained robot postures are necessary. Additionally, autonomous navigation requires planners that can identify energy-optimal paths based on a global map understanding, which the current study does not address.
However, the loss models developed in this research provide a valuable foundation for future challenges in energy optimization for legged robots, potentially extending their range for applications such as inspecting disaster zones, carrying deliveries, or other uses.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
