科技科学

A quadruped robot designed to complete a marathon on a single battery charge

Nature作者：Choongin Lee2026年9月23日正文已收录本站

Data availability

All data presented in the paper are available at Zenodo (https://doi.org/10.5281/zenodo.14825866)61.

Code availability

Code for locomotion policy training and reward-ablation evaluation is available at GitHub (https://github.com/railabatkaist/raisimGym_nature).

References

  1. Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V. & Hutter, M. Learning quadrupedal locomotion over challenging terrain. Sci. Robot. 5, eabc5986 (2020).

    Article  PubMed  Google Scholar 

  2. McGeer, T. Dynamic walking robots and the w prize [industry/research news]. IEEE Robot. Autom. Mag. 14, 13–15 (2007).

    Article  Google Scholar 

  3. Shin, Y.-H. et al. Design of KAIST HOUND, a quadruped robot platform for fast and efficient locomotion with mixed-integer nonlinear optimization of a gear train. In Proc. 2022 International Conference on Robotics and Automation (ICRA) 6614–6620 (IEEE, 2022).

  4. Bjelonic, F. et al. Learning-based design and control for quadrupedal robots with parallel-elastic actuators. IEEE Robot. Autom. Lett. 8, 1611–1618 (2023).

    Article  Google Scholar 

  5. Fu, Z., Kumar, A., Malik, J. & Pathak, D. Minimizing energy consumption leads to the emergence of gaits in legged robots. In Proc. 5th Conference on Robot Learning (eds Faust, A., Hsu, D. & Neumann, G.) 928–937 (PMLR, 2022).

  6. Yang, Y., Zhang, T., Coumans, E., Tan, J. & Boots, B. Fast and efficient locomotion via learned gait transitions. In Proc. 5th Conference on Robot Learning (eds Faust, A., Hsu, D. & Neumann, G.) 773–783 (PMLR, 2022).

  7. RaiLab KAIST. (1x speed) raibo2 completes full marathon (42.195km) in 4:19:52—world’s first robot marathon record. YouTube https://youtu.be/-HMFoa3g9nA (2024).

  8. Park, H.-W., Wensing, P. M. & Kim, S. High-speed bounding with the MIT Cheetah 2: control design and experiments. Int. J. Robot. Res. 36, 167–192 (2017).

    Article  Google Scholar 

  9. Jin, Y., Liu, X., Shao, Y., Wang, H. & Yang, W. High-speed quadrupedal locomotion by imitation-relaxation reinforcement learning. Nat. Mach. Intell. 4, 1198–1208 (2022).

    Article  Google Scholar 

  10. Ji, G., Mun, J., Kim, H. & Hwangbo, J. Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion. IEEE Robot. Autom. Lett. 7, 4630–4637 (2022).

    Article  Google Scholar 

  11. Shin, Y.-H. et al. Reinforcement learning for high-speed quadrupedal locomotion with motor operating region constraints: mitigating motor model discrepancies through torque clipping in realistic motor operating region. IEEE Robot. Autom. Mag. 32, 49–59 (2024).

    Article  Google Scholar 

  12. Choi, S. et al. Learning quadrupedal locomotion on deformable terrain. Sci. Robot. 8, eade2256 (2023).

    Article  PubMed  Google Scholar 

  13. Miki, T. et al. Learning robust perceptive locomotion for quadrupedal robots in the wild. Sci. Robot. 7, eabk2822 (2022).

    Article  PubMed  Google Scholar 

  14. Kim, H. et al. High-speed control and navigation for quadrupedal robots on complex and discrete terrain. Sci. Robot. 10, eads6192 (2025).

    Article  PubMed  Google Scholar 

  15. Unitree B2: go beyond the limits. Unitree Robotics https://www.unitree.com/b2 (2023).

  16. Wensing, P. M. et al. Proprioceptive actuator design in the MIT Cheetah: impact mitigation and high-bandwidth physical interaction for dynamic legged robots. IEEE Trans. Robot. 33, 509–522 (2017).

    Article  Google Scholar 

  17. Seok, S. et al. Design principles for energy-efficient legged locomotion and implementation on the MIT Cheetah robot. IEEE/ASME Trans. Mechatronics 20, 1117–1129 (2015).

    Article  Google Scholar 

  18. Kitano, S., Hirose, S., Horigome, A. & Endo, G. TITAN-XIII: sprawling-type quadruped robot with ability of fast and energy-efficient walking. ROBOMECH J. 3, 1–16 (2016).

    Article  Google Scholar 

  19. Valsecchi, G. et al. Barry: a high-payload and agile quadruped robot. IEEE Robot. Autom. Lett. 8, 6939–6946 (2023).

    Article  Google Scholar 

  20. Chen, L. et al. Optimized foothold planning and posture searching for energy-efficient quadruped locomotion over challenging terrains. In Proc. 2020 IEEE International Conference on Robotics and Automation (ICRA) 399–405 (IEEE, 2020).

  21. 22nd Sangju Gotgam Marathon record system. SPCT Co. https://time.spct.kr/m2.php?EVENT_NO=20241117017&TargetYear=2024&currentPage=1&BIB_NO=042004 (2024).

  22. Tucker, V. A. The energetic cost of moving about: walking and running are extremely inefficient forms of locomotion. Much greater efficiency is achieved by birds, fish—and bicyclists. Am. Sci. 63, 413–419 (1975).

    CAS  PubMed  ADS  Google Scholar 

  23. Bledt, G. et al. MIT Cheetah 3: design and control of a robust, dynamic quadruped robot. In Proc. 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2245–2252 (IEEE, 2018).

  24. Katz, B., Di Carlo, J. & Kim, S. Mini Cheetah: a platform for pushing the limits of dynamic quadruped control. In Proc. 2019 International Conference on Robotics and Automation (ICRA) 6295–6301 (IEEE, 2019).

  25. Chang, L., Jeng, Y.-R. & Huang, P.-Y. Modeling and analysis of the meshing losses of involute spur gears in high-speed and high-load conditions. J. Tribol. 135, 011504 (2013).

    Article  Google Scholar 

  26. Margaria, R. Positive and negative work performances and their efficiencies in human locomotion. Int. Z. Angew. Physiol. 25, 339–351 (1968).

    CAS  PubMed  Google Scholar 

  27. Hutter, M. et al. Starleth: a compliant quadruped robot for fast, efficient, and versatile locomotion. In Adaptive Mobile Robotics (eds Azad, A. K. M., Cowan, N. J., Tokhi, M. O., Virk, G. S. & Eastman, R. D.) 483–490 (World Scientific, 2012).

  28. Ding, L. et al. Brief analysis of a Big-Dog quadruped robot. Chin. Mech. Eng. 23, 505–514 (2012).

    Google Scholar 

  29. Li, M. et al. Leg prototype of a bio-inspired quadruped robot. Robot 36, 21–28 (2014).

    CAS  Google Scholar 

  30. Shi, Y., Zhang, M. & Zhang, X. Design and analysis of quadruped robot’s leg with flexible joints. J. Huazhong Univ. Sci. Technol. (Nat. Sci. Ed.) 45, 17–22 (2017).

    Google Scholar 

  31. Mi, C., Slemon, G. & Bonert, R. Modeling of iron losses of permanent-magnet synchronous motors. IEEE Trans. Ind. Appl. 39, 734–742 (2003).

    Article  Google Scholar 

  32. Lakkas, G. MOSFET power losses and how they affect power-supply efficiency. Analog Applications J. 1Q, SLYT664 (Texas Instruments, 2016).

  33. Gopalan, A. Calculating Power Dissipation for a H-bridge or Half Bridge Driver Technical Report SLVA504A (Texas Instruments, 2021).

  34. Bobbert, A. C. Energy expenditure in level and grade walking. J. Appl. Physiol. 15, 1015–1021 (1960).

    Article  Google Scholar 

  35. Tucker, V. A. Energetic cost of locomotion in animals. Comp. Biochem. Physiol. 34, 841–846 (1970).

    Article  CAS  PubMed  Google Scholar 

  36. EV Database. Compare electric vehicles https://ev-database.org/ (EV Database.org, 2025).

  37. Fuelly. Track and compare your mpg https://www.fuelly.com/car (Fuelly, 2026).

  38. McGeer, T. Passive dynamic walking. Int. J. Robot. Res. 9, 62–82 (1990).

    Article  Google Scholar 

  39. Bhounsule, P. et al. Low-bandwidth reflex-based control for lower power walking: 65 km on a single battery charge. Int. J. Robot. Res. 33, 1305–1321 (2014).

    Article  Google Scholar 

  40. Li, Q., Liu, G., Tang, J. & Zhang, J. A simple 2D straight-leg passive dynamic walking model without foot-scuffing problem. In Proc. 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 5155–5161 (IEEE, 2016).

  41. Lee, J. et al. Learning robust autonomous navigation and locomotion for wheeled-legged robots. Sci. Robot. 9, eadi9641 (2024).

    Article  PubMed  Google Scholar 

  42. Hutter, M. et al. Anymal—toward legged robots for harsh environments. Adv. Robot. 31, 918–931 (2017).

    Article  Google Scholar 

  43. Boston Dynamics. Spot - the agile mobile robot. Boston Dynamics https://bostondynamics.com/products/spot/ (2019).

  44. Unitree B1: Land Overlord. Unitree https://www.unitree.com/b1 (2022).

  45. Meet ANYmal your new inspector. ANYbotics https://www.anybotics.com/robotics/anymal/ (2021).

  46. Vision 60: Agile and autonomous quadruped. Ghost Robotics https://www.ghostrobotics.io/vision-60 (2021).

  47. X30: Industrial application quadruped robot. DEEP Robotics https://deeprobotics.cn/en/index/product3.html (2023).

  48. Hartley, R., Ghaffari, M., Eustice, R. M. & Grizzle, J. W. Contact-aided invariant extended Kalman filtering for robot state estimation. Int. J. Robot. Res. 39, 402–430 (2020).

    Article  Google Scholar 

  49. Raibert, M. H. & Tello, E. R. Legged robots that balance. IEEE Expert 1, 89 (1986).

    Article  ADS  Google Scholar 

  50. American Gear Manufacturers Association. Fundamental Rating Factors and Calculation Methods for Involute Spur and Helical Gear Teeth. ANSI/AGMA 2101-D04 (American Gear Manufacturers Association, 2004).

  51. Infineon Technologies AG. EiceDRIVER high voltage gate driver IC evaluation board. Application Note EVAL_6EDL04N02PR, Rev. 1.0 (Infineon Technologies AG, 2013).

  52. Schulman, J., Wolski, F., Dhariwal, P., Radford, A. & Klimov, O. Proximal policy optimization algorithms. Preprint at https://arxiv.org/abs/1707.06347 (2017).

  53. Pinto, L., Andrychowicz, M., Welinder, P., Zaremba, W. & Abbeel, P. Asymmetric actor critic for image-based robot learning. Preprint at https://arxiv.org/abs/1710.06542 (2017).

  54. Hwangbo, J., Lee, J. & Hutter, M. Per-contact iteration method for solving contact dynamics. IEEE Robot. Autom. Lett. 3, 895–902 (2018).

    Article  Google Scholar 

  55. Bengio, Y., Louradour, J., Collobert, R. & Weston, J. Curriculum learning. In Proc. 26th Annual International Conference on Machine Learning (eds Bottou, L. & Littman, M.) 41–48 (ACM, 2009).

  56. Mysore, S., Mabsout, B., Mancuso, R. & Saenko, K. Regularizing action policies for smooth control with reinforcement learning. In Proc. 2021 IEEE International Conference on Robotics and Automation (ICRA) 1810–1816 (IEEE, 2021).

  57. Peng, X. B., Andrychowicz, M., Zaremba, W. & Abbeel, P. Sim-to-real transfer of robotic control with dynamics randomization. In Proc. 2018 IEEE International Conference on Robotics and Automation (ICRA) 3803–3810 (IEEE, 2018).

  58. EtherCAT – the Ethernet fieldbus. EtherCAT Technology Group https://www.ethercat.org/default.htm (2024).

  59. Tong, W., Wu, S., Sun, J. & Zhu, L. Iron loss analysis of permanent magnet synchronous motor with an amorphous stator core. In Proc. 2016 IEEE Vehicle Power and Propulsion Conference (VPPC) 1–6 (IEEE, 2016).

  60. Bertotti, G. General properties of power losses in soft ferromagnetic materials. IEEE Trans. Magn. 24, 621–630 (1988).

    Article  ADS  Google Scholar 

  61. Lee, C. Dataset for ‘RAIBO2: highly efficient quadruped robot completing full marathon with a single battery charge’. Zenodo https://doi.org/10.5281/zenodo.14825866 (2025).

Download references

Funding

This work was supported by the Samsung Research Funding and Incubation Center of Samsung Electronics under Project Number SRFC-IT2002-02.

Author information

Author notes

  1. These authors contributed equally: Choongin Lee, Donghoon Youm, Jeongsoo Park

Authors and Affiliations

  1. Robotics and Artificial Intelligence Laboratory, KAIST, Daejeon, Korea

    Choongin Lee, Donghoon Youm, Jeongsoo Park, Jeonghyun Lee, Suyoung Choi, Gwanghyeon Ji, Juhyeok Mun, Moonkyu Jung, Hyungho Chris Choi, Hyeongjun Kim, Hyunsik Oh, Kyungmin Nam, Minho Lee, Jungwoo Hur, Donghyuk Choi, Donggeon Kim, Yeonsu An & Jemin Hwangbo

Authors

  1. Choongin Lee
  2. Donghoon Youm
  3. Jeongsoo Park
  4. Jeonghyun Lee
  5. Suyoung Choi
  6. Gwanghyeon Ji
  7. Juhyeok Mun
  8. Moonkyu Jung
  9. Hyungho Chris Choi
  10. Hyeongjun Kim
  11. Hyunsik Oh
  12. Kyungmin Nam
  13. Minho Lee
  14. Jungwoo Hur
  15. Donghyuk Choi
  16. Donggeon Kim
  17. Yeonsu An
  18. Jemin Hwangbo

Contributions

C.L., D.Y., J.P. and J. Hwangbo designed the research and analysed the data. C.L. designed the electrical system. D.Y. trained the control policy. J.P. designed the mechanical hardware. C.L., J.L., S.C., G.J., J.M. and J. Hwangbo developed the robot software. C.L., D.Y., J.P., J.L., S.C., G.J., J.M., M.J., H.C., H.K., H.O., K.N., M.L., J. Hur, D.C., D.K. and Y.A. assisted in designing the experimental setup and supporting the marathon performance evaluation. C.L., D.Y., J.P., J.L. and J. Hwangbo co-wrote the paper. All authors gave final approval for publication.

Corresponding author

Correspondence to Jemin Hwangbo.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature thanks Pulkit Agrawal and Donghyun Kim for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Fig. 1 Design implementations for leg mechanism and motor driver circuit.

a, Lightweight leg mechanism and actuator structure. The hip mount and calf link were mass-reduced through stress-analysis-guided material removal. Needle and crossed roller bearings are used to maintain structural rigidity with compact geometry. b, Motor driver architecture showing the integrated communication processor, power-control stage, MOSFET arrangement, and multiple current-sensing configurations (Hall and shunt).

Extended Data Fig. 2 Systematic approach to motor driver loss reduction through sequential design improvements.

a,b, Experimental setup for loss measurements: complete test bench configuration (a) and integrated motor driver-stator system (b). c-h, Comprehensive loss characterization across design parameters: current sensing method comparison (c), MOSFET conduction loss analysis at low voltage operation (20 V) (d), MOSFET switching loss analysis at high voltage operation (80 V) (e), PCB copper thickness effects (f), gate resistance modification (g), and switching frequency impact (h).

Extended Data Fig. 3 Sensitivity of total loss and cost of transport (COT) to hardware design parameters.

Total loss and COT were evaluated across locomotion speeds by applying systematic hardware perturbations in simulation, using realistic locomotion profiles obtained from the RL controller. Actuator-level loss components (motor and driver power) were grounded on experimentally measured data. An asterisk (*) marks the baseline configuration used in RAIBO2, while all other curves represent comparison conditions used for sensitivity analysis. a,b, Gear-ratio scaling; c,d, Motor models; e,f, Body-weight variations; g,h, Calf-link mass (filled vs. lightened); i,j, Current-sensing methods (Hall 1.5 mΩ, shunt values); k,l, Gate-resistor values, comparing RAIBO2 (10 Ω*) with the Infineon 6EDL04N02PR EVM (180 Ω).

Extended Data Fig. 4 Reinforcement learning framework and terrain classification for locomotion policy development.

a, Detailed system architecture showing agent networks (privileged information, proprioceptive observation, actor-critic networks) and environment elements (simulator, rewards, curriculum). b, Training environments classified into three terrain types based on maximum foot contact angle: slopes and hills (Type 1), standard stairs and steps (Type 2), and stairs with nosing and pipes (Type 3).

Extended Data Fig. 5 Experimental validation of locomotion policy for efficiency improvement.

a,b, RAIBO2 running at controlled speed on treadmill for efficiency measurements (Supplementary Video 2 and 3). c,d, Comparison of foot velocity profiles between proposed and baseline controllers, with detailed view of contact moment (d). e, Pre-contact velocity distribution histogram across different control strategies. f, Knee joint torque profiles demonstrating the effect of actuator Joule loss reward. g, Torque distribution histogram comparing policies with and without actuator Joule loss reward consideration.

Extended Data Table 1 RAIBO2 robot and motor specifications

Full size table

Extended Data Table 2 Network parameters and training hyperparameters

Full size table

Extended Data Table 3 Reward functions and ablation study

Full size table

Extended Data Table 4 Driver ablation study experimental conditions and MOSFET specifications

Full size table

Supplementary information

Supplementary Data 1 (download XLSX )

Comparative energy and COT data across robots, animals and vehicles. Compiled data on mass, energy consumption and capacity, locomotion speed, range and COT for quadrupedal robots, wheeled robots, animals, electric vehicles, gasoline vehicles and diesel vehicles. Values are obtained or estimated from the cited literature and publicly available sources for cross-system comparison.

Supplementary Video 1 (download MP4 )

Versatile locomotion capabilities of RAIBO2 in various outdoor environments. A quadrupedal robot RAIBO2 demonstrates robust locomotion across diverse real-world scenarios. The video shows the robot maintaining stable gait patterns while running alongside human marathoners, ascending stairs, executing high-speed locomotion on flat terrain and navigating through snowy conditions, highlighting its adaptability to challenging environmental conditions.

Supplementary Video 2 (download MP4 )

Comparative analysis of collision reward impact on RAIBO2 running efficiency. Experimental comparison of RAIBO2’s running performance with and without collision reward in the control policy at controlled speed (4 m s−1) on a treadmill. Real-time waveform visualization demonstrates total loss, electric loss and mechanical loss metrics. The video clearly illustrates substantially higher mechanical losses in the control policy without collision reward, validating the effectiveness of the collision reward term.

Supplementary Video 3 (download MP4 )

Quantitative analysis of body height effects on RAIBO2 running efficiency. Systematic comparison of RAIBO2’s running performance at three different body heights (0.45 m, 0.50 m, 0.55 m) at controlled speed (4 m s−1) on a treadmill. Real-time waveform visualization shows total loss, electric loss and mechanical loss metrics. The video demonstrates the inverse relationship between body height and electric loss, with higher configurations resulting in reduced electric losses during locomotion.

Supplementary Video 4 (download MP4 )

RAIBO2 Marathon Achievement Documentation. Complete documentation of RAIBO2’s marathon performance including course overview and running highlights. The video presents key segments of the 42.195-km marathon course and showcases sustained stable locomotion throughout the event, demonstrating the robot’s endurance and reliability.

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Lee, C., Youm, D., Park, J. et al. A quadruped robot designed to complete a marathon on a single battery charge. Nature (2026). https://doi.org/10.1038/s41586-026-11102-5

Download citation

  • Received:

  • Accepted:

  • Published:

  • Version of record:

  • DOI: https://doi.org/10.1038/s41586-026-11102-5