Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

KLEIYN : A Quadruped Robot with an Active Waist for Both Locomotion and Wall Climbing

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A quadruped robot with a single bending torso joint climbs narrow walls at about fifty times the speed of previous climbing robots.

desk verdict First quadruped chimney climb with an active waist is a real proof of concept, but the speed claims rest on un-repeated demos and the waist benefit is sim-only. read the letter →

arxiv 2507.06562 v2 pith:N6C472UW submitted 2025-07-09 cs.RO

Keita Yoneda , Kento Kawaharazuka , Temma Suzuki , Takahiro Hattori , Kei Okada This is my paper
classification cs.RO
keywords quadrupedrobotwallclimbingchimneywaistjointreinforcementlearningcurriculumsim-to-realtransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper seeks to show that a quadruped robot equipped with one extra joint in its torso can both walk on the ground and climb up between two vertical walls, using a reinforcement-learning policy trained entirely in simulation. The robot, KLEIYN, braces its feet against opposing walls to climb, a mode called chimney climbing, and reached real walls of 800 to 1000 mm width at an average speed near 150 mm/s. That speed is roughly fifty times the 3 mm/s reported for the only previous chimney-climbing robot. The authors also argue that the waist joint is what makes narrow-wall climbing possible and improves tracking of a commanded climb speed.

What carries the argument

Two mechanisms carry the result. The first is the active waist joint, a pitch-axis 1-DOF motor connecting the front and back body links, which lets KLEIYN bend its torso to brace in narrow chimneys and store and release energy during the stance-to-swing transition. The second is Contact-Guided Curriculum Learning (CGCL), a training scheme that starts the robot on a floor joined to the wall by an elliptical curve with horizontal radius r=0.3 m and gradually reduces r to zero, so the agent experiences bracing contact early and learns vertical climbing only after the contact pattern is established.

What would settle it

Run the same trained policy on a 3 m high, 900 mm wide real wall and record whether the robot sustains the claimed 150 mm/s average speed to the top; the paper notes motors overheat and trigger thermal shutdowns during climbing, so a sustained climb that stops well short of 3 m would falsify the practical speed claim.

Watch

Extended reading notes

Core claim

KLEIYN is an 18 kg quadruped with four three-joint legs and a 1-DOF pitch-axis waist, giving 13 degrees of freedom in total. Using Proximal Policy Optimization in the Isaac Gym simulator, with an Asymmetric Actor-Critic setup and a new curriculum called Contact-Guided Curriculum Learning (CGCL) that gradually morphs the floor-wall junction from a curved U-shape into a vertical surface, the robot learned to climb by alternately bracing against two walls and lifting its legs. In the real world, the learned policy climbed plywood walls of 800, 900, and 1000 mm width, reaching a maximum height of 1.0 m in the 800 mm case at an average speed of 170 mm/s. Removing the waist joint in simulation lowered the tracking score of the commanded velocity, with the gap growing on narrower walls, and a 750 mm wall could only be climbed with the waist enabled even though the body is 760 mm long. The paper concludes that the active waist both widens the range of climbable wall widths and improves climbing performance by letting the robot use recoil when lifting its legs.

Load-bearing premise

The central claim depends on the simulated training environment being faithful enough to transfer to the real robot, since the simplified torque model assumes even load distribution across the four legs, ignores leg mass, fixes friction at 0.8, and only randomizes friction, mass, and inertia within modest ranges.

Editorial extensions

If this is right

  • A single quadruped with one torso joint can perform both flat-ground locomotion and vertical chimney climbing without grippers or other climbing-specialized hardware.
  • The waist joint lets a 760 mm long robot climb a 750 mm wide wall in simulation, so the active torso genuinely extends the range of passable gaps.
  • CGCL has a threshold around r=0.25 m: with junctions more vertical than that from the start, the policy fails to learn climbing within 20,000 iterations, showing the curriculum is doing real work.
  • The policy generalizes to wall widths never seen in training (750 mm in simulation; 800 and 1000 mm in reality), suggesting the learned bracing strategy is not tuned to one geometry.
  • Real-world climbing speed around 150 to 170 mm/s is about 50 times faster than the 3 mm/s reported for the prior chimney-climbing robot SiLVIA.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to close the loop with the onboard LiDAR: the robot fell when it drifted to the wall's edge because localization was only recorded, not used in control; feeding wall-edge distance into the observation would likely prevent those falls.
  • The waist joint's recoil benefit could transfer to other contact-rich quadruped skills such as jumping onto ledges, stepping over gaps, or climbing stairs with overhangs, since the mechanism only adds one actuated degree of freedom.
  • If the torque and friction mismatch that caused the 1050 mm failure were closed by modeling motor saturation or richer foot friction in simulation, the same policy could plausibly climb wider walls and sustain longer climbs than the 1.0 m demonstrated.
  • CGCL is a generic idea: any task where the agent must discover an uncomfortable contact pattern before it can progress, such as sit-to-stand, wall push-offs, or grasping, could use the same smooth-to-vertical terrain morphing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents KLEIYN, an 18 kg quadruped robot with 13 DOF including an active pitch-axis waist joint, and claims it can both walk on flat ground and climb narrow vertical gaps (chimney climbing) using an RL policy trained in Isaac Gym with a proposed curriculum, Contact-Guided Curriculum Learning (CGCL). The authors report real-world climbing on walls of 800, 900, and 1000 mm width, with a headline average speed of 150 mm/s (50x faster than the cited SiLVIA hexapod), and a simulation ablation showing that the waist joint improves tracking performance on narrow walls. A separate locomotion policy is demonstrated outdoors on steps and uneven stones.

Significance. If substantiated, this work would be a meaningful advance in legged locomotion: it presents, to my knowledge, the first quadruped robot that combines a waist joint with RL-based chimney climbing, and the CGCL curriculum is a simple and potentially reusable idea for learning bracing motions. The real-world climbing demonstration is a genuine proof of concept, and the CGCL ablation in Sec. V-A.2 is clear and informative. However, the quantitative support for the headline speed claim is incomplete, the waist-joint benefit is not validated on hardware, and the sim-to-real transfer evidence is weakened by undocumented single trials and acknowledged failures. With additional experiments and more rigorous reporting, the contribution could be significant.

major comments (4)
  1. [Abstract; Sec. V-B] The abstract claims that 'KLEIYN successfully climbed walls ranging from 800 mm to 1000 mm in width at an average speed of 150 mm/s', but Sec. V-B reports only a single average speed of 170 mm/s for the 800 mm case and gives no speed values, trial counts, success criteria, or variance for the 900 mm and 1000 mm cases. The 150 mm/s figure in the abstract is therefore unsupported by the data presented. Please report per-wall-width average speeds, the number of repeated trials, success rates, and a clear definition of a successful climb, or revise the claim.
  2. [Sec. V-B] The real-world climbing results appear to be single demonstrations, with no explicit repetition count, success rate, or statistical spread. Given that the authors document the robot failing on a 1050 mm wall, slipping frequently during stance, falling after reaching the wall edge, and experiencing motor overheating, the reliability of the 800-1000 mm success is not established without repeated trials and quantitative success metrics. Please add a repetition protocol and report the result of each trial (e.g., height reached, fall events, slip counts).
  3. [Sec. V-A.3] The claim that the waist joint improves climbing performance, especially on narrow walls, is based solely on simulation and is not tested on the real robot with the waist locked. Since this is one of the four stated contributions and is highlighted in the abstract and conclusion, please add a hardware comparison with the waist joint locked or explicitly qualify the claim as simulation-only.
  4. [Sec. IV-B.4; Sec. V-A.3] The tracking score used to compare waist/no-waist policies is the same exponential function f(x,0.01) that appears as the 'tracking velocity' reward term in Table II. The comparison therefore partly measures how well each policy optimized its own training objective, which can bias the evaluation. Reporting an independent metric (e.g., mean absolute vertical velocity error) would make the waist-joint claim more convincing and should be added.
minor comments (5)
  1. [Sec. V-B] The text says the robot 'frequently' slipped during stance, but no quantitative slip rate or slip magnitude is given; please quantify this observation to support the claim of robustness to disturbances.
  2. [Abstract; Sec. II-A; Sec. V-B] The phrase '50 times faster than conventional robots' is imprecise: the only comparator cited is the hexapod SiLVIA (3 mm/s). Please specify the baseline and its conditions, and clarify that this is a comparison to a specific prior platform rather than a general class of robots.
  3. [Sec. III-C] The discrepancy between the torque analysis (which predicts ~20 Nm sufficient) and the real-world failure at 1050 mm is explained only qualitatively. Reporting measured joint torques during climbing and comparing them to the model predictions would strengthen the design validation.
  4. [Fig. 10] The caption mentions a maximum climbing height of 1.0 m, but the text does not state the wall height used in the real-world setup or whether the robot reached the top; please state these details explicitly.
  5. [Sec. IV-A] The control loop is described as 50 Hz, but the motor drivers run internal position control at an unspecified rate; please specify the motor control frequency and any latency to clarify the control architecture.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the climbing claim is a direct real-world demonstration, and the reward-derived tracking metric is applied symmetrically to both ablated models.

full rationale

The paper's central derivation chain is self-contained and does not reduce to its own inputs. The real-world chimney-climbing result (Sec. V-B) is a direct physical demonstration on 800, 900, and 1000 mm walls, not a quantity computed from a fitted model. The 50x speed comparison uses the externally reported SiLVIA speed of 3 mm/s from references [13], [14], which are independent prior works, not self-citations. The motor torque assessment (Sec. III-C) is a forward static analysis using a Jacobian model with stated assumptions (even load distribution, friction 0.8, neglected leg mass); it informs hardware selection but is not used to generate the climbing prediction. The RL training uses standard PPO in Isaac Gym with externally cited components, and the CGCL curriculum is a training procedure, not a hidden constraint that forces the climbing outcome. The only mildly self-referential element is the waist/no-waist comparison in Sec. V-A.3, where the tracking score is the mean of f(x, 0.01), the same exponential function used in the tracking velocity reward term. However, both the with-waist and without-waist models are trained with the same reward, and the metric directly measures the policies' achieved tracking behavior; this is an evaluation of the training objective, not a fitted parameter renamed as a prediction or an equation that reduces by construction. The paper also reports real-world failures and discrepancies (1050 mm wall failure, slipping, edge falls, motor overheating), which are correctness and generalization risks rather than circularity. The self-citation to MEVIUS [19] for leg design is not load-bearing to the core climbing claim. Overall, no circular step can be exhibited from the text.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper contributes hardware and an empirical policy; it relies on standard RL transfer assumptions and a simplified static model. The hand-tuned reward weights, curriculum schedule, and domain-randomization ranges are the main free parameters. No new physical entities are postulated.

free parameters (4)
  • CGCL curve parameter r = starts at 0.3 m, reduced to 0; learning fails below r=0.25
    Hand-chosen curriculum schedule; threshold at r=0.25 found empirically in Sec. V-A.2.
  • Reward weights (Table II) = e.g., tracking velocity 3.0, termination -500, orientation -10.0
    Hand-tuned scalar weights define the learned behavior; no sensitivity analysis is given.
  • Domain randomization ranges = friction 0.7-0.95; velocity perturbations 0-1 m/s; unspecified mass/inertia variation
    Chosen by the authors to improve sim-to-real transfer; exact values are stated only partially in Sec. IV-B.2.
  • Training command velocity range = 0.0 to 0.6 m/s
    Sampled uniformly during training and not tied to any external requirement.
assumptions (5)
  • standard math Torque is computed by the static Jacobian relation tau = -J^T f_ext (Eq. 1) with no dynamics.
    This is a standard quasi-static link between endpoint force and joint torque.
  • domain assumption Bracing load is evenly distributed across all four legs and leg mass is neglected.
    Sec. III-C; used to size motors. Real load distribution is dynamic and uneven.
  • domain assumption Contact friction coefficient is 0.8 for motor sizing and 0.7-0.95 in training.
    Sec. III-C and IV-B.2; the actual friction on plywood is not measured.
  • domain assumption Isaac Gym simulation with domain randomization is a sufficient proxy for real contact, friction, and motor torque for zero-shot policy transfer.
    Sec. IV-B and V-B; the paper itself reports discrepancies (torque insufficiency at 1050 mm, larger downward velocity, motor overheating).
  • ad hoc to paper The tracking score used to compare waist and no-waist policies is the same exponential function f(x,0.01) as the tracking reward.
    Sec. V-A.3; this is a paper-specific evaluation choice that ties the metric to the training objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KLEIYN : A Quadruped Robot with an Active Waist for Both Locomotion and Wall Climbing." pith.science (2026). https://pith.science/paper/N6C472UW

@misc{pith2026250706562,
  author       = {Pith},
  title        = {Pith review of: KLEIYN : A Quadruped Robot with an Active Waist for Both Locomotion and Wall Climbing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N6C472UW}},
  note         = {Machine review of arXiv:2507.06562}
}
read the original abstract

In recent years, advancements in hardware have enabled quadruped robots to operate with high power and speed, while robust locomotion control using reinforcement learning (RL) has also been realized. As a result, expectations are rising for the automation of tasks such as material transport and exploration in unknown environments. However, autonomous locomotion in rough terrains with significant height variations requires vertical movement, and robots capable of performing such movements stably, along with their control methods, have not yet been fully established. In this study, we developed the quadruped robot KLEIYN, which features a waist joint, and aimed to expand quadruped locomotion by enabling chimney climbing through RL. To facilitate the learning of vertical motion, we introduced Contact-Guided Curriculum Learning (CGCL). As a result, KLEIYN successfully climbed walls ranging from 800 mm to 1000 mm in width at an average speed of 150 mm/s, 50 times faster than conventional robots. Furthermore, we demonstrated that the introduction of a waist joint improves climbing performance, particularly enhancing tracking ability on narrow walls.

Figures

Figures reproduced from arXiv: 2507.06562 by the authors.

Figure 1
Figure 1. Overview of KLEIYN, the quadruped robot with an active waist [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Design of KLEIYN’s waist joint. The motor output is reduced by [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 5
Figure 5. The training and real-world control framework of KLEIYN. [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Since this study aims to achieve both horizontal and [PITH_FULL_IMAGE:figures/full_fig_p003_6.png]
Figure 6
Figure 6. Figure 6: The terrain used for wall climbing learning. Initially, the floor [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 9
Figure 9. Figure 9: Comparison of tracking performance with and without the [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 8
Figure 8. Figure 8: Comparison of models trained for 20,000 iterations with varying [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 10
Figure 10. Figure 10: Wall-climbing experiments on walls with a width of 800 mm, 900 mm, and 1000 mm. The figure shows the climbing motion and the corresponding [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: Outdoor locomotion experiments. The robot successfully traversed [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. WARL: Wrench-Augmented Reinforcement Learning for Task-Agnostic Learning in Legged Robots

    cs.RO 2026-07 conditional novelty 4.0 of 10

    Adding a simulated torso wrench during early RL training and gradually removing it lets a quadruped learn six locomotion tasks with a shared reward, yielding a joint-only policy in simulation.

Reference graph

Works this paper leans on

22 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hutter, C

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V . Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, and M. Hoepflinger, “ANYmal - a highly

    work page

  2. [2]

    Mini Cheetah: A Platform for Pushing the Limits of Dynamic Quadruped Control,

    B. Katz, J. D. Carlo, and S. Kim, “Mini Cheetah: A Platform for Pushing the Limits of Dynamic Quadruped Control,” inProceedings of the 2019 IEEE International Conference on Robotics and Automation, 2019, pp. 6295–6301

    work page 2019

  3. [3]

    Go1 (Unitree Robotics),

    “Go1 (Unitree Robotics),” https://m.unitree.com/go1

    work page

  4. [4]

    Robust quadruped jumping via deep reinforcement learning,

    G. Bellegarda, C. Nguyen, and Q. Nguyen, “Robust quadruped jumping via deep reinforcement learning,”Robotics and Autonomous Systems, vol. 182, p. 104799, 2024

    2024

  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, 2019. mobile and dynamic quadrupedal robot,” inProceedings of the 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2016, pp. 38–44

    work page 2019

  6. [6]

    Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning,

    V . Tsounis, M. Alge, J. Lee, F. Farshidian, and M. Hutter, “Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3699–3706, 2020

    work page 2020

  7. [7]

    Controlling the solo12 quadruped robot with deep rein- forcement learning,

    M. Aractingi, P.-A. L ´eziart, T. Flayols, J. Perez, T. Silander, and P. Sou`eres, “Controlling the solo12 quadruped robot with deep rein- forcement learning,”scientific Reports, vol. 13, no. 1, p. 11945, 2023

    work page 2023

  8. [8]

    Learning vision-based bipedal locomotion for challenging terrain,

    H. Duan, B. Pandit, M. S. Gadde, B. Van Marum, J. Dao, C. Kim, and A. Fern, “Learning vision-based bipedal locomotion for challenging terrain,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 56–62

    2024

Show all 22 references
  1. [9]

    Learning robust perceptive locomotion for quadrupedal robots in the wild,

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning robust perceptive locomotion for quadrupedal robots in the wild,”Science Robotics, vol. 7, no. 62, 2022

    2022

  2. [10]

    Lemur 3: A limbed climbing robot for extreme terrain mobility in space,

    A. Parness, N. Abcouwer, C. Fuller, N. Wiltsie, J. Nash, and B. Kennedy, “Lemur 3: A limbed climbing robot for extreme terrain mobility in space,” in2017 IEEE International Conference on Robotics and Automation (ICRA), 2017, pp. 5467–5473

    2017

  3. [11]

    Scaler: A tough versatile quadruped free-climber robot,

    Y . Tanaka, Y . Shirai, X. Lin, A. Schperberg, H. Kato, A. Swerdlow, N. Kumagai, and D. Hong, “Scaler: A tough versatile quadruped free-climber robot,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 5632–5639

    2022

  4. [12]

    LORIS: A lightweight free-climbing robot for extreme terrain exploration,

    P. Nadan, S. Backus, and A. M. Johnson, “LORIS: A lightweight free-climbing robot for extreme terrain exploration,” inIEEE Intl. Conference on Robotics and Automation, 2024, pp. 18 480–18 486

    2024

  5. [13]

    Multi-limbed robot vertical two wall climbing based on static indeterminacy modeling and feasibility region analysis,

    X. Lin, H. Krishnan, Y . Su, and D. W. Hong, “Multi-limbed robot vertical two wall climbing based on static indeterminacy modeling and feasibility region analysis,” in2018 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS), 2018

    2018

  6. [14]

    Optimiza- tion based motion planning for multi-limbed vertical climbing robots,

    X. Lin, J. Zhang, J. Shen, G. Fernandez, and D. W. Hong, “Optimiza- tion based motion planning for multi-limbed vertical climbing robots,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019, pp. 1918–1925

    2019

  7. [15]

    Learning to walk in minutes using massively parallel deep reinforcement learning,

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter, “Learning to walk in minutes using massively parallel deep reinforcement learning,” in Proceedings of the 2022 Conference on Robot Learning, 2022, pp. 91–100

    2022

  8. [16]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,”Science Robotics, vol. 5, no. 47, 2020

    2020

  9. [17]

    Asymmetric actor critic for image-based robot learning,

    L. Pinto, M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel, “Asymmetric actor critic for image-based robot learning,”arXiv preprint arXiv:1710.06542, 2017

    2017 arXiv

  10. [18]

    Learn- ing humanoid locomotion over challenging terrain,

    I. Radosavovic, S. Kamat, T. Darrell, and J. Malik, “Learn- ing humanoid locomotion over challenging terrain,”arXiv preprint arXiv:2410.03654, 2024

    2024 arXiv

  11. [19]

    MEVIUS: A Quadruped Robot Easily Constructed through E-Commerce with Sheet Metal Welding and Ma- chining,

    K. Kawaharazuka, S. Inoue, T. Suzuki, S. Yuzaki, S. Sawaguchi, K. Okada, and M. Inaba, “MEVIUS: A Quadruped Robot Easily Constructed through E-Commerce with Sheet Metal Welding and Ma- chining,” in2024 IEEE-RAS International Conference on Humanoid Robots, 2024

    2024

  12. [20]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017. [Online]. Available: https://arxiv.org/abs/1707.06347

    2017 arXiv

  13. [21]

    Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter,

    W. Xu and F. Zhang, “Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3317–3324, 2021

    2021

  14. [22]

    Gpu-accelerated robotic simulation for distributed reinforcement learning,

    J. Liang, V . Makoviychuk, A. Handa, N. Chentanez, M. Macklin, and D. Fox, “Gpu-accelerated robotic simulation for distributed reinforcement learning,” 2018. [Online]. Available: https://arxiv.org/ abs/1810.05762

    2018 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.