pith. sign in

arxiv: 2605.27909 · v1 · pith:H2G5JDULnew · submitted 2026-05-27 · 💻 cs.RO

S-Cheetah: A Novel Quadrupedal Robot with a 3-DOF Active Spine Learning Agile Locomotion

Zimu Li , Weibang Bai This is my paper

Pith reviewed 2026-06-29 12:13 UTC · model grok-4.3

classification 💻 cs.RO
keywords quadrupedal robotactive spine3-DOF spinereinforcement learningagile locomotiongallop gaitaerial self-rightingbiomimetic design
0
0 comments X

The pith

A quadrupedal robot with a 3-DOF active spine reaches 6.9 m/s and develops aerial self-righting.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces S-Cheetah, a quadrupedal robot whose spine can flex and rotate along three axes in the same way biological spines do. It pairs this hardware with a reinforcement learning setup that uses curriculum acceleration and three targeted rewards to force the spine into active use during movement. Experiments then record a peak running speed of 6.9 m/s, an in-place turn rate of 7.2 rad/s, and an unplanned ability to right itself in mid-air so that it lands on all four feet. The central argument is that adding full tri-axial spinal motion, when the learning process is shaped to exploit it, measurably widens the range of agile behaviors a quadruped can perform. A reader would care because current robots with simpler spines fall short of the maneuverability seen in animals that use their backs extensively.

Core claim

S-Cheetah is built with a 3-DOF bio-inspired serial active spine that performs sagittal flexion/extension, lateral bending, and axial rotation. A specialized reinforcement learning framework combines an acceleration curriculum with a gallop gait reward, a spine undulation reward, and a spine steering reward to drive active spine participation. Under this training the robot produces a rotary G2 gallop that reaches 6.9 m/s, an in-place turning rate of 7.2 rad/s, and an emergent feline-style aerial self-righting behavior that allows stable four-footed landings from arbitrary free-fall orientations. Evaluations across multiple tasks show the 3-DOF spine improves overall locomotive agility.

What carries the argument

The 3-DOF bio-inspired serial active spine, whose sagittal, lateral, and axial rotations are actively recruited by a reinforcement learning controller that uses curriculum learning plus gallop, undulation, and steering rewards.

If this is right

  • The rotary G2 gallop gait becomes usable at 6.9 m/s once the spine is actively engaged.
  • In-place turning rate reaches 7.2 rad/s through coordinated spine steering.
  • An aerial self-righting reflex emerges that stabilizes landings from any initial orientation.
  • Locomotive agility increases across a range of tasks when the full three-axis spine is available and rewarded.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar multi-axis spines paired with reward-shaped learning could improve balance recovery on uneven ground.
  • The self-righting behavior may reduce the need for separate orientation sensors or pre-landing corrections.
  • Testing the same spine on robots of different sizes or masses could show whether the agility gains scale linearly.
  • Other legged platforms might adopt three-axis spinal modules if the reward design is reused without major retuning.

Load-bearing premise

The tailored rewards and curriculum in the reinforcement learning framework are what cause the spine to engage in ways that produce the measured speed, turning, and self-righting gains.

What would settle it

An otherwise identical robot without the 3-DOF spine, or trained with the same framework but without the spine-specific rewards, fails to reach 6.9 m/s, 7.2 rad/s turning, or the aerial self-righting behavior.

Figures

Figures reproduced from arXiv: 2605.27909 by Weibang Bai, Zimu Li.

Figure 1
Figure 1. Figure 1: S-Cheetah : A 3-DOF Spine-driven Cheetah-inspired Quadrupedal Robot for Agile Locomotion. Our proposed S-Cheetah demonstrates agile locomotion capabilities across diverse scenarios. (a) High-speed rotary G2 galloping. (b) Maneuverable in-place turning and moving turning. (c) Emergent feline-inspired aerial self-righting. Abstract— The biological spine of quadrupeds enables sagit￾tal flexion/extension, late… view at source ↗
Figure 4
Figure 4. Figure 4: Basic framework of the reinforcement learning [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Design of S-Cheetah’s spine joints and their corre [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rotary G2 gallop performance and its gait diagrams. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: The ωz tracking performance during turning [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Angular variations of the three spine joints during [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Evaluation of the “figure-8” path-tracking performance. (a) Spatial trajectories of the active-spine and rigid-trunk [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Variations in roll angle and roll rate during aerial [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Angular variations of the three spine joints during [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
read the original abstract

The biological spine of quadrupeds enables sagittal flexion/extension, lateral bending, and axial rotation, playing a crucial role in highly agile and dexterous locomotion. While numerous studies have integrated active spinal joints into quadrupedal robots to enhance agility, most designs simplify control complexity by reducing spinal degrees of freedom (DOF), failing to achieve the spatial tri-axial rotation characteristic of biological spines. Consequently, replicating a multi-DOF biomimetic spine and effectively leveraging it to empower the agile locomotion of quadrupedal robots remains a significant research challenge. In this study, we present S-Cheetah, a quadrupedal robot featuring a 3-DOF bio-inspired serial active spine capable of biomimetic spatial tri-axial rotation. To empower the robot to fully utilize this active spine, we developed a specialized reinforcement learning framework to actively promote the engagement of the introduced spine and maximize the robot's locomotive capabilities by integrating an acceleration curriculum learning strategy with tailored reward functions, such as a gallop gait reward, a spine undulation reward, and a spine steering reward. Experimental results demonstrate that S-Cheetah can achieve a peak speed of 6.9 m/s using the rotary G2 gallop gait and an in-place turning rate of 7.2 rad/s. Besides, the system exhibits an emergent, feline-inspired aerial self-righting capability, allowing it to land stably on four feet from arbitrary orientations during free fall. Finally, through extensive evaluations across diverse locomotion tasks, we prove that the introduction of the proposed 3-DOF spine comprehensively enhances the locomotive agility of quadrupedal robots. Project website: himmy-robotics.github.io/scheetah

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The paper introduces S-Cheetah, a quadrupedal robot equipped with a 3-DOF bio-inspired serial active spine enabling sagittal flexion/extension, lateral bending, and axial rotation. It proposes a specialized reinforcement learning framework that combines an acceleration curriculum learning strategy with tailored reward functions (gallop gait reward, spine undulation reward, and spine steering reward) to promote spine engagement. Experimental results are claimed to show a peak speed of 6.9 m/s via rotary G2 gallop, an in-place turning rate of 7.2 rad/s, an emergent feline-inspired aerial self-righting behavior, and overall enhancement of locomotive agility through the 3-DOF spine across diverse tasks.

Significance. If substantiated, the demonstration of a full tri-axial active spine yielding measurable agility gains (speed and turning) plus emergent self-righting would provide concrete evidence for the value of biomimetic multi-DOF spinal actuation in quadrupedal robots, extending beyond reduced-DOF designs common in prior work.

major comments (3)
  1. [Abstract] Abstract: Headline performance claims (peak speed of 6.9 m/s and turning rate of 7.2 rad/s) are stated without any reference to experimental protocol, measurement apparatus, number of trials, environmental conditions, or statistical measures, rendering it impossible to evaluate whether the data support the central claims about spine-enabled agility.
  2. [Abstract (RL framework description)] The specialized RL framework is described at a high level (acceleration curriculum plus gallop gait, spine undulation, and spine steering rewards) but supplies no equations, hyperparameter values, or ablation results showing that these components actively promote 3-DOF spine engagement rather than being incidental to the policy.
  3. [Abstract (final sentence)] No baseline comparisons (e.g., against 1-DOF or 2-DOF spine variants or state-of-the-art quadrupeds without active spines) or verification steps are mentioned, so the claim that the 3-DOF spine 'comprehensively enhances' agility cannot be assessed for load-bearing contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each of the three major comments below, focusing on revisions to the abstract to improve clarity and substantiation of the claims while respecting length constraints.

read point-by-point responses

  1. Referee: [Abstract] Abstract: Headline performance claims (peak speed of 6.9 m/s and turning rate of 7.2 rad/s) are stated without any reference to experimental protocol, measurement apparatus, number of trials, environmental conditions, or statistical measures, rendering it impossible to evaluate whether the data support the central claims about spine-enabled agility.

    Authors: We agree that the abstract would benefit from additional context on validation. The full manuscript (Section V) details the use of an OptiTrack motion-capture system for speed and turning-rate measurements, conducted over multiple trials in a controlled indoor setting with flat terrain. We will revise the abstract to include a brief clause noting that peak values were obtained via repeated experiments with motion-capture verification. revision: yes

  2. Referee: [Abstract (RL framework description)] The specialized RL framework is described at a high level (acceleration curriculum plus gallop gait, spine undulation, and spine steering rewards) but supplies no equations, hyperparameter values, or ablation results showing that these components actively promote 3-DOF spine engagement rather than being incidental to the policy.

    Authors: Equations for the three reward terms appear in Section IV-B, with hyperparameters listed in the appendix and ablation results in Section VI-C confirming their role in driving spine motion. Abstract length precludes full equations, but we will expand the description to explicitly state that the curriculum and rewards are designed to promote 3-DOF engagement and direct readers to the main text for equations and ablations. revision: partial

  3. Referee: [Abstract (final sentence)] No baseline comparisons (e.g., against 1-DOF or 2-DOF spine variants or state-of-the-art quadrupeds without active spines) or verification steps are mentioned, so the claim that the 3-DOF spine 'comprehensively enhances' agility cannot be assessed for load-bearing contribution.

    Authors: Section VI presents quantitative comparisons against 1-DOF/2-DOF spine ablations and other quadruped platforms, isolating the contribution of the 3-DOF spine. We will revise the final sentence of the abstract to note that the agility gains are substantiated by these comparative evaluations across locomotion tasks. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical demonstration only

full rationale

The paper is an empirical robotics demonstration describing hardware (3-DOF spine robot) and an RL training setup with custom rewards. No equations, first-principles derivations, fitted parameters presented as predictions, or self-citation chains appear in the abstract or referenced full text. Claims rest on experimental outcomes (speed, turning rate, self-righting) rather than any reduction of outputs to inputs by construction. No load-bearing steps match the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical model, free parameters, axioms, or invented physical entities are introduced in the abstract; the contribution is a hardware design plus empirical RL training.

pith-pipeline@v0.9.1-grok · 5832 in / 1170 out tokens · 48930 ms · 2026-06-29T12:13:01.453238+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 7 canonical work pages · 1 internal anchor

  1. [1]

    R. M. Alexander,Principles of animal locomotion. Princeton university press, 2003

    2003

  2. [2]

    The quadrupedal gaits of vertebrates,

    M. Hildebrand, “The quadrupedal gaits of vertebrates,”Bioscience, vol. 39, no. 11, p. 766, 1989

    1989

  3. [3]

    Dynamical determinants enabling two different types of flight in cheetah gallop to enhance speed through spine movement,

    T. Kamimura, S. Aoi, Y . Higurashi, N. Wada, K. Tsuchiya, and F. Matsuno, “Dynamical determinants enabling two different types of flight in cheetah gallop to enhance speed through spine movement,” Scientific Reports, vol. 11, no. 1, p. 9631, 2021

    2021

  4. [4]

    Mini cheetah: A platform for push- ing the limits of dynamic quadruped control,

    B. Katz, J. D. Carlo, and S. Kim, “Mini cheetah: A platform for push- ing the limits of dynamic quadruped control,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 6295– 6301

    2019

  5. [5]

    Learning agile and dynamic motor skills for legged robots,

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter, “Learning agile and dynamic motor skills for legged robots,”Science Robotics, vol. 4, no. 26, p. eaau5872,

  6. [6]

    Available: https://www.science.org/doi/abs/10.1126/sc irobotics.aau5872

    [Online]. Available: https://www.science.org/doi/abs/10.1126/sc irobotics.aau5872

    work page doi:10.1126/sc

  7. [7]

    Anymal parkour: Learning agile navigation for quadrupedal robots,

    D. Hoeller, N. Rudin, D. Sako, and M. Hutter, “Anymal parkour: Learning agile navigation for quadrupedal robots,”Science Robotics, vol. 9, no. 88, p. eadi7566, 2024. [Online]. Available: https://www.science.org/doi/abs/10.1126/scirobotics.adi7566

    work page doi:10.1126/scirobotics.adi7566 2024

  8. [8]

    High-speed control and navigation for quadrupedal robots on complex and discrete terrain,

    H. Kim, H. Oh, J. Park, Y . Kim, D. Youm, M. Jung, M. Lee, and J. Hwangbo, “High-speed control and navigation for quadrupedal robots on complex and discrete terrain,”Science Robotics, vol. 10, no. 102, p. eads6192, 2025. [Online]. Available: https://www.science. org/doi/abs/10.1126/scirobotics.ads6192

    work page doi:10.1126/scirobotics.ads6192 2025

  9. [9]

    (2025, jul) Chinese-made robotic dog hits 10.3 m/s, rivals human sprinters

    Xinhua News Agency. (2025, jul) Chinese-made robotic dog hits 10.3 m/s, rivals human sprinters. Xinhua. [Online; accessed 9-January- 2026]. [Online]. Available: http://english.news.cn/20250707/cc729a6 a50854206ae621fbe142886d5/c.html

    work page arXiv 2025

  10. [10]

    (2025, may) Fastest 100 m by a quadrupedal robot

    Guinness World Records. (2025, may) Fastest 100 m by a quadrupedal robot. Guinness World Records. [Online; accessed 9-January-2026]. [Online]. Available: https://www.guinnessworldrecords.com/world-r ecords/625586-fastest-100-m-by-a-quadrupedal-robot

    2025

  11. [11]

    Learning active spine behaviors for dynamic and efficient locomotion in quadruped robots,

    S. Bhattacharya, A. Singla, Abhimanyu, D. Dholakiya, S. Bhatnagar, B. Amrutur, A. Ghosal, and S. Kolathaya, “Learning active spine behaviors for dynamic and efficient locomotion in quadruped robots,” in2019 28th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN), 2019, pp. 1–6

    2019

  12. [12]

    Twisting spine or rigid torso: Exploring quadrupedal morphology via trajectory optimization,

    J. Caporale, Z. Feng, S. Rozen-Levy, A. M. Carter, and D. E. Koditschek, “Twisting spine or rigid torso: Exploring quadrupedal morphology via trajectory optimization,” in2023 IEEE International Conference on Robotics and Automation, no. 886, 2023

    2023

  13. [13]

    A robust quadruped robot with twisting waist for flexible motions,

    Q. Qian, X. Wei, Z. Zhang, J. Tu, Y . Zhang, T. Hou, X. Gao, P. Zhai, and L. Zhang, “A robust quadruped robot with twisting waist for flexible motions,”arXiv preprint arXiv:2410.05884, 2024

    work page arXiv 2024

  14. [14]

    Kleiyn : A quadruped robot with an active waist for both locomotion and wall climbing,

    K. Yoneda, K. Kawaharazuka, T. Suzuki, T. Hattori, and K. Okada, “Kleiyn : A quadruped robot with an active waist for both locomotion and wall climbing,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, pp. 8783–8789

    2025

  15. [15]

    Enhancing the flexibility of a quadruped robot with a 2- dof active spine using nonlinear model predictive control,

    Z. Yang, Z. Xu, H. Rong, S. Mo, Y . Chen, Z. Chen, T. Wang, and H. Cheng, “Enhancing the flexibility of a quadruped robot with a 2- dof active spine using nonlinear model predictive control,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, pp. 12 526–12 532

    2025

  16. [16]

    Spined torso renders advanced mobility for quadrupedal locomotion,

    J. Wang, J. Cheng, J. Hu, W. Gao, and S. Zhang, “Spined torso renders advanced mobility for quadrupedal locomotion,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 1457–1463

    2024

  17. [17]

    A Generic Continuous Multi-Joint Spinal Robotic System for Agile and Accurate Behaviors with GNN-MPC method,

    Y . Wu, Z. Zhou, R. Liu, L. Zheng, and H. Cheng, “A Generic Continuous Multi-Joint Spinal Robotic System for Agile and Accurate Behaviors with GNN-MPC method,” inProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025

    2025

  18. [18]

    Predictive control of a spined quadrupedal robot based on a dual rigid-body model,

    J. Hu, W. Gao, X. Zhang, J. Cheng, and S. Zhang, “Predictive control of a spined quadrupedal robot based on a dual rigid-body model,” in2024 IEEE International Conference on Robotics and Biomimetics (ROBIO), 2024, pp. 292–298

    2024

  19. [19]

    Dynamic locomotion of a quadruped robot with active spine via model predictive control,

    W. Li, Z. Zhou, and H. Cheng, “Dynamic locomotion of a quadruped robot with active spine via model predictive control,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 1185–1191

    2023

  20. [20]

    Running, jumping, hunting, and scavenging: Functional analysis of vertebral mobility and backbone properties in carnivorans,

    R. I. Belyaev, P. Nikolskaia, A. V . Bushuev, A. A. Panyutina, D. A. Kozhanova, and N. E. Prilepskaya, “Running, jumping, hunting, and scavenging: Functional analysis of vertebral mobility and backbone properties in carnivorans,”Journal of Anatomy, vol. 244, no. 2, pp. 205–231, 2024

    2024

  21. [21]

    A dynamical explanation of the falling cat phenomenon,

    T. R. Kane and M. Scher, “A dynamical explanation of the falling cat phenomenon,”International journal of solids and structures, vol. 5, no. 7, pp. 663–670, 1969

    1969

  22. [22]

    Rsl-rl: A learning library for robotics research,

    C. Schwarke, M. Mittal, N. Rudin, D. Hoeller, and M. Hutter, “Rsl-rl: A learning library for robotics research,”arXiv preprint arXiv:2509.10771, 2025

    work page arXiv 2025

  23. [23]

    Proximal Policy Optimization Algorithms

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  24. [24]

    16 ways to gallop: Ener- getics and body dynamics of high-speed quadrupedal gaits,

    Y . G. Alqaham, J. Cheng, and Z. Gan, “16 ways to gallop: Ener- getics and body dynamics of high-speed quadrupedal gaits,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025, pp. 16 168–16 174

    2025

  25. [25]

    robot lab: Rl extension library for robots, based on isaaclab

    Z. Fan, “robot lab: Rl extension library for robots, based on isaaclab.”

  26. [26]

    Available: https://github.com/fan-ziqi/robot lab

    [Online]. Available: https://github.com/fan-ziqi/robot lab