2026-09-24
KAIST's RAIBO2 pairs hardware with an RL locomotion policy and finishes a marathon in 4:19:52 on one charge, COT 0.25 versus 0.37 for humans.
Quadrupeds spend energy at the joints just to hold body weight, and they dump kinetic energy at each footfall, so their range per charge stays far behind wheeled platforms. Prior work usually attacked one loss term in hardware or in control. KAIST’s Robotics and Artificial Intelligence Lab put the whole energy-loss model on one ledger and built RAIBO2: force-transparent lightweight mechanics, a low-resistance motor driver, and a low-dissipation locomotion policy.
The Nature body is paywalled. Method details below come from the abstract, figure captions, Extended Data blurbs, and the paper’s own code and data links. Training hyperparameters, the full ablation tables, and the marathon time-series numbers in the main text were not retrieved.
Public materials split efficiency into three tracks. Mechanics stress force transparency and mass. The driver stresses a low-resistance power stage. The policy is a learned locomotion controller: actor-critic, with proprioception and privileged information, trained in Raisim under a curriculum. Terrains are grouped by maximum foot contact angle into slopes and hills, ordinary stairs, and stairs with nosing and pipes.
Extended Data also names two energy-facing rewards: actuator Joule loss and collision. Treadmill tests compare foot velocity, contact timing, and knee-torque histograms against a baseline controller. Hardware ablations cover PCB copper thickness, gate resistance (10 Ω on RAIBO2 versus 180 Ω on an Infineon evaluation board), switching frequency, and current sensing. In simulation, hardware parameters are perturbed on gait profiles taken from the RL controller, and total loss and cost of transport (COT, energy per unit weight per unit distance) are read out.
Network sizes, PPO hyperparameters, and reward weights live in Extended Data Tables 2–3. The landing page gives the table titles, not the cell values. Those tables should not be treated as read.
The headline result in the abstract is a real marathon: 42.195 km on one battery in 4 hours 19 minutes 52 seconds, COT 0.25 versus the paper’s human benchmark of 0.37. Range per charge is more than three times that of existing quadrupeds. The loss-analysis figure is labeled with 1,447 Wh. Comparative data sit on Zenodo, including the Fig. 1 spreadsheet across robots, animals, and vehicles; per-platform numbers printed in the main text remain behind the paywall.
Extended Data promises reward and driver ablations. A supplementary video is explicitly about the collision reward on the treadmill. Without the full paper, the size of those effects cannot be quoted.
Legged-robot demos often stop at “it runs and it climbs.” This paper treats endurance as an engineering number that can be compared with human COT, and it refuses to locate the answer in a bigger network alone. The loss model, the power stage, and the policy rewards move together. For people doing sim-to-real legged control, the portable pieces are energy-facing rewards such as Joule loss and collision, plus the habit of scoring hardware perturbations on real policy rollouts.
Code and data are public: policy training in the raisimGymnature GitHub repo, logs on Zenodo. Reproduction does not have to wait on the paywall.
The main text was not obtained, so reward weights, network scale, the identity of the baseline controller, and race-day wind, surface, and pack temperature cannot be checked. The abstract’s “more than three times the travel range” does not say which production quadrupeds sit in the denominator, or whether speed and payload are matched.
COT 0.25 versus 0.37 is transport cost, not pace. A 4 hour 20 minute marathon is a finishable amateur human time and about twice the world record. Reading “lower COT than a human” as “outruns a human” is a mistake. Hardware and policy contributions are not separated in the public materials, so the marathon should not be booked as a pure RL win.