arxiv: 2605.09944 · v1 · submitted 2026-05-11 · 💻 cs.RO

Recognition: no theorem link

Explicit Stair Geometry Conditioning for Robust Humanoid Locomotion

Jianguo Zhang , Wentai Xu , Shusheng Ye , Yuxiang He , Weimin Qi , Qinbo Sun , Ning Ding , Liguang Zhou

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:27 UTC · model grok-4.3

classification 💻 cs.RO

keywords humanoid locomotionstair climbingreinforcement learningexplicit conditioningPPOterrain adaptationUnitree G1

0 comments

The pith

Explicit stair geometry parameters condition a PPO policy for robust humanoid climbing on varying stairs.

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

The paper establishes that replacing implicit terrain encodings or blind proprioception with direct conditioning on compact stair parameters lets a locomotion policy anticipate and adjust to geometric changes. It extracts step height, step depth, and yaw angle to modulate swing-foot clearance and stride in a PPO policy. Simulation results show better generalization to stair heights outside the training set. Real-world tests on the Unitree G1 humanoid confirm reliable indoor and outdoor traversals, including ascent of 33 consecutive steps without failure. This matters because it directly targets the sensitivity to height variations and perception uncertainty that limit current learning-based humanoid locomotion.

Core claim

Extracting a compact set of interpretable geometric parameters—step height, step depth, and current yaw angle relative to the robot heading—and using them to condition a Proximal Policy Optimization locomotion policy enables proactive modulation of swing-foot clearance and stride characteristics according to stair structure.

What carries the argument

Explicit stair geometry conditioning framework that supplies compact parameters directly to the policy instead of high-dimensional latent features.

If this is right

The conditioned policy generalizes to stair heights beyond those seen in training.
The Unitree G1 humanoid performs reliable stair traversal in both indoor and outdoor environments.
The robot completes ascent of 33 consecutive outdoor steps without failure.

Where Pith is reading between the lines

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

The approach may extend to other discontinuous surfaces if analogous low-dimensional geometric descriptors can be defined and sensed.
Success depends on perception modules that remain accurate during dynamic motion, suggesting integration with improved depth or visual estimators.
It could reduce the dimensionality needed in terrain encoders for broader locomotion tasks.

Load-bearing premise

The compact set of stair geometry parameters can be accurately extracted and supplied to the policy in real time from the robot's sensors under varying lighting, occlusion, and motion.

What would settle it

Demonstrating repeated failures or falls on stairs whose heights lie outside the training distribution, or in settings where sensor data prevents reliable extraction of the three parameters, would show the claimed generalization and robustness do not hold.

Figures

Figures reproduced from arXiv: 2605.09944 by Jianguo Zhang, Liguang Zhou, Ning Ding, Qinbo Sun, Shusheng Ye, Weimin Qi, Wentai Xu, Yuxiang He.

**Figure 1.** Figure 1: Explicit stair geometry conditioning for robust humanoid locomotion. The local stair structure is parameterized by step height hstep, step depth dstep, and current yaw angle θ current yaw defined relative to the robot heading direction. These explicit geometric parameters directly condition the locomotion policy, enabling anticipatory and adaptive gait modulation across varying stair configurations. A key… view at source ↗

**Figure 2.** Figure 2: Overview of the proposed explicit stair geometry conditioning framework. During training, a teacher branch leverages [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Simulation sequence of humanoid stair climbing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: MuJoCo simulation results of stair climbing with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Ablation study on stair perception representations [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Time-sequence key frames from real humanoid robot experiments in an indoor environment (left to right). The robot [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

**Figure 7.** Figure 7: Time-sequence key frames from real humanoid robot outdoor step-climbing experiments (1 s–8 s). The robot [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗

read the original abstract

Robust humanoid stair climbing remains challenging due to geometric discontinuities, sensitivity to step height variations, and perception uncertainty in real-world environments. Existing learning-based locomotion policies often rely on implicit terrain representations or blind proprioceptive feedback, limiting their ability to generalize across varying stair geometries and to anticipate required gait adjustments. This paper proposes an explicit stair geometry conditioning framework for robust humanoid stair climbing. Instead of encoding terrain as high-dimensional latent features, we extract a compact set of interpretable geometric parameters, including step height, step depth, and current yaw angle relative to the robot heading. These explicit stair parameters directly condition a Proximal Policy Optimization (PPO)-based locomotion policy, enabling proactive modulation of swing-foot clearance and stride characteristics according to stair structure. Simulation experiments demonstrate improved generalization across unseen stair heights beyond the training distribution. Real-world experiments on the Unitree G1 humanoid validate reliable indoor and outdoor stair traversal. In challenging outdoor scenarios, the robot successfully ascends 33 consecutive steps without failure, demonstrating robustness and practical deployability.

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.

Desk Editor's Note private letter to a colleague

The paper conditions a PPO policy on three explicit stair parameters for better sim generalization and real G1 hardware runs, but the perception accuracy that carries the claim is not measured.

read the letter

Here's the core of it: the authors extract step height, depth, and yaw angle from the terrain and feed those three numbers directly into a PPO locomotion policy so the robot can adjust swing clearance and stride before it reaches the next step. This is a clear shift from implicit latent encodings or pure proprioception, and the simulation results indicate the policy handles stair heights outside the training distribution. On the Unitree G1 the hardware tests show reliable indoor and outdoor traversal, including one outdoor sequence of 33 consecutive steps without failure.

Referee Report

3 major / 1 minor

Summary. The paper proposes an explicit stair geometry conditioning framework for robust humanoid stair climbing. A compact set of interpretable parameters (step height, step depth, and current yaw angle) is extracted and used to directly condition a PPO-based locomotion policy, enabling proactive adjustments to swing-foot clearance and stride. Simulation experiments claim improved generalization to unseen stair heights, while real-world tests on the Unitree G1 report reliable indoor/outdoor traversal, including successful ascent of 33 consecutive outdoor steps.

Significance. If the explicit low-dimensional conditioning can be shown to deliver measurable gains over implicit or proprioceptive baselines and if the real-time geometry extraction remains accurate under realistic conditions, the approach would represent a practical advance in interpretable humanoid control for discontinuous terrains, potentially improving deployability without requiring high-dimensional latent encodings.

major comments (3)

[Abstract] Abstract: The claim of 'improved generalization across unseen stair heights beyond the training distribution' is unsupported by any quantitative metrics, success rates, traversal statistics, error bars, or baseline comparisons, preventing assessment of whether the explicit conditioning provides a substantive benefit.
[Abstract] Abstract: The real-world result of '33 consecutive steps without failure' in challenging outdoor scenarios rests on the unverified assumption that stair geometry parameters can be extracted accurately in real time; no error metrics, sensor fusion details, failure-mode analysis, or robustness tests under lighting/occlusion variations are supplied, leaving the load-bearing link between explicit inputs and observed robustness unverified.
[Abstract] Abstract: No ablation studies isolating the contribution of explicit conditioning versus standard proprioceptive policies, nor comparisons to implicit terrain representations, are reported, which is required to substantiate the central methodological claim.

minor comments (1)

[Abstract] The abstract would benefit from specifying the ranges of stair heights/depths tested in simulation and the sensor modalities used for geometry extraction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation for major revision. We address each point below and will make the necessary revisions to strengthen the quantitative support and methodological clarity in the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The claim of 'improved generalization across unseen stair heights beyond the training distribution' is unsupported by any quantitative metrics, success rates, traversal statistics, error bars, or baseline comparisons, preventing assessment of whether the explicit conditioning provides a substantive benefit.

Authors: We agree that the abstract would benefit from including specific quantitative evidence. The simulation results section of the manuscript reports success rates, traversal statistics, and comparisons to proprioceptive baselines with error bars across multiple seeds. To make these benefits clear from the abstract, we will revise the abstract to incorporate key quantitative metrics and baseline comparisons from the experiments. revision: yes
Referee: [Abstract] Abstract: The real-world result of '33 consecutive steps without failure' in challenging outdoor scenarios rests on the unverified assumption that stair geometry parameters can be extracted accurately in real time; no error metrics, sensor fusion details, failure-mode analysis, or robustness tests under lighting/occlusion variations are supplied, leaving the load-bearing link between explicit inputs and observed robustness unverified.

Authors: The manuscript describes the real-time stair geometry extraction pipeline in the methods section, relying on depth sensing and onboard computation. We acknowledge that additional validation of this component is warranted. In the revised manuscript, we will include error metrics for the extracted parameters, details on the sensor fusion approach, failure-mode analysis, and robustness evaluations under different lighting and occlusion conditions. revision: yes
Referee: [Abstract] Abstract: No ablation studies isolating the contribution of explicit conditioning versus standard proprioceptive policies, nor comparisons to implicit terrain representations, are reported, which is required to substantiate the central methodological claim.

Authors: We recognize the importance of ablations to demonstrate the specific advantage of explicit conditioning. The current experiments include comparisons to a standard proprioceptive policy, but we will add dedicated ablation studies and comparisons against implicit terrain encoding methods in the revised version to better isolate and substantiate the contribution of the proposed approach. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method with independent validation

full rationale

The paper describes an explicit stair geometry conditioning approach for a PPO locomotion policy, using extracted parameters (step height, depth, yaw) to modulate gait. No equations, derivations, or self-citations are present that reduce claimed generalization or real-world performance to quantities fitted from the same data or defined by construction. Simulation and hardware results are reported as separate experiments, with the central claims resting on empirical outcomes rather than tautological reductions. The extraction step is an engineering assumption but does not create circularity in any derivation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the ability to extract accurate geometric parameters in real time and on standard RL training assumptions; no new physical entities are postulated.

free parameters (2)

PPO training hyperparameters
Standard reinforcement-learning training constants that are tuned to produce the policy.
Stair geometry extraction thresholds or models
Parameters used to compute the compact set of height, depth, and yaw from sensor data.

axioms (1)

domain assumption Stair geometry parameters can be accurately and timely perceived from onboard sensors.
Invoked as the foundation for direct policy conditioning.

pith-pipeline@v0.9.0 · 5494 in / 1313 out tokens · 59209 ms · 2026-05-12T04:27:47.348809+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 4 internal anchors

[1]

Ali Aalipour and Alireza Khani. Data-driven h-infinity control with a real-time and efficient reinforcement learning algorithm: An appli- cation to autonomous mobility-on-demand systems.arXiv preprint arXiv:2309.08880, 2023

work page arXiv 2023
[2]

Legged locomotion in challenging terrains using egocentric vision

Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In Conference on robot learning, pages 403–415. PMLR, 2023

work page 2023
[3]

Legs as manipu- lator: Pushing quadrupedal agility beyond locomotion

Xuxin Cheng, Ashish Kumar, and Deepak Pathak. Legs as manipu- lator: Pushing quadrupedal agility beyond locomotion. In2023 IEEE International Conference on Robotics and Automation (ICRA), 2023

work page 2023
[4]

Extreme parkour with legged robots

Xuxin Cheng, Kexin Shi, Ananye Agarwal, and Deepak Pathak. Extreme parkour with legged robots. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11443–11450. IEEE, 2024

work page 2024
[5]

Learning to walk in the real world with minimal human effort

Sehoon Ha, Peng Xu, Zhenyu Tan, Sergey Levine, and Jie Tan. Learning to walk in the real world with minimal human effort. In Jens Kober, Fabio Ramos, and Claire Tomlin, editors,Proceedings of the 2020 Conference on Robot Learning, volume 155 ofProceedings of Machine Learning Research, pages 1110–1120. PMLR, 16–18 Nov 2021

work page 2020
[6]

Learning to walk via deep reinforcement learning

Tuomas Haarnoja, Sehoon Ha, Aurick Zhou, Jie Tan, George Tucker, and Sergey Levine. Learning to walk via deep reinforcement learning. InProceedings of Robotics: Science and Systems, FreiburgimBreisgau, Germany, June 2019

work page 2019
[7]

Hwangbo, J

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, 2019

work page 2019
[8]

Dynamic locomotion on slippery ground.IEEE Robotics and Automation Letters, 4(4):4170–4176, 2019

Fabian Jenelten, Jemin Hwangbo, Fabian Tresoldi, C Dario Bellicoso, and Marco Hutter. Dynamic locomotion on slippery ground.IEEE Robotics and Automation Letters, 4(4):4170–4176, 2019

work page 2019
[9]

Not only rewards but also constraints: Applications on legged robot locomotion.IEEE Transactions on Robotics, 40:2984–3003, 2024

Yunho Kim, Hyunsik Oh, Jeonghyun Lee, Jinhyeok Choi, Gwanghyeon Ji, Moonkyu Jung, Donghoon Youm, and Jemin Hwangbo. Not only rewards but also constraints: Applications on legged robot locomotion.IEEE Transactions on Robotics, 40:2984–3003, 2024

work page 2024
[10]

Learning quadrupedal locomotion over challenging terrain.Science robotics, 5(47):eabc5986, 2020

Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain.Science robotics, 5(47):eabc5986, 2020

work page 2020
[11]

Learning agile bipedal mo- tions on a quadrupedal robot

Yunfei Li, Jinhan Li, Wei Fu, and Yi Wu. Learning agile bipedal mo- tions on a quadrupedal robot. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 9735–9742. IEEE, 2024

work page 2024
[12]

Hybrid internal model: Learning agile legged loco- motion with simulated robot response

Junfeng Long, ZiRui Wang, Quanyi Li, Liu Cao, Jiawei Gao, and Jiangmiao Pang. Hybrid internal model: Learning agile legged loco- motion with simulated robot response. InThe Twelfth International Conference on Learning Representations, 2024

work page 2024
[13]

Multi-agent actor-critic for mixed cooperative- competitive environments.Advances in neural information processing systems, 30, 2017

Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative- competitive environments.Advances in neural information processing systems, 30, 2017

work page 2017
[14]

Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning

Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

work page internal anchor Pith review arXiv 2021
[15]

Walk these ways: Tuning robot control for generalization with multiplicity of behavior

Gabriel B Margolis and Pulkit Agrawal. Walk these ways: Tuning robot control for generalization with multiplicity of behavior. In Conference on Robot Learning, pages 22–31. PMLR, 2023

work page 2023
[16]

Rapid locomotion via reinforcement learning.The Interna- tional Journal of Robotics Research, 43(4):572–587, 2024

Gabriel B Margolis, Ge Yang, Kartik Paigwar, Tao Chen, and Pulkit Agrawal. Rapid locomotion via reinforcement learning.The Interna- tional Journal of Robotics Research, 43(4):572–587, 2024

work page 2024
[17]

Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning

Mayank Mittal, Pascal Roth, James Tigue, Antoine Richard, Octi Zhang, Peter Du, Antonio Serrano-Mu˜noz, Xinjie Yao, Ren´e Zurbr¨ugg, Nikita Rudin, Lukasz Wawrzyniak, Milad Rakhsha, Alain Denzler, Eric Heiden, Ales Borovicka, Ossama Ahmed, Iretiayo Akinola, Abrar Anwar, Mark T. Carlson, Ji Yuan Feng, Animesh Garg, Renato Gasoto, Lionel Gulich, Yijie Guo, M...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[18]

Robust reinforcement learning.Neural computation, 17(2):335–359, 2005

Jun Morimoto and Kenji Doya. Robust reinforcement learning.Neural computation, 17(2):335–359, 2005

work page 2005
[19]

Dreamwaq: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning

I Made Aswin Nahrendra, Byeongho Yu, and Hyun Myung. Dreamwaq: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5078–5084. IEEE, 2023

work page 2023
[20]

Robust quadrupedal locomotion on sloped terrains: A linear policy approach

Kartik Paigwar, Lokesh Krishna, Shalabh Bhatnagar, Ashitava Ghosal, Bharadwaj Amrutur, Shishir Kolathaya, et al. Robust quadrupedal locomotion on sloped terrains: A linear policy approach. InConference on Robot Learning, pages 2257–2267. PMLR, 2021

work page 2021
[21]

Sim-to-real transfer of robotic control with dynamics ran- domization

Xue Bin Peng, Marcin Andrychowicz, Wojciech Zaremba, and Pieter Abbeel. Sim-to-real transfer of robotic control with dynamics ran- domization. In2018 IEEE international conference on robotics and automation (ICRA), pages 3803–3810. IEEE, 2018

work page 2018
[22]

Dynamic walking with compliance on a cassie bipedal robot

Jacob Reher, Wen-Loong Ma, and Aaron D Ames. Dynamic walking with compliance on a cassie bipedal robot. In2019 18th European Control Conference (ECC), pages 2589–2595. IEEE, 2019

work page 2019
[23]

Learn- ing to walk in minutes using massively parallel deep reinforcement learning

Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learn- ing to walk in minutes using massively parallel deep reinforcement learning. InConference on robot learning, pages 91–100. PMLR, 2022

work page 2022
[24]

High-Dimensional Continuous Control Using Generalized Advantage Estimation

John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation.arXiv preprint arXiv:1506.02438, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015
[25]

Proximal Policy Optimization Algorithms

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017
[26]

Chase Kew, Tianyu Li, Linda Luu, Xue Bin Peng, Sehoon Ha, Jie Tan, and Sergey Levine

Laura M Smith, J. Chase Kew, Tianyu Li, Linda Luu, Xue Bin Peng, Sehoon Ha, Jie Tan, and Sergey Levine. Learning and Adapting Agile Locomotion Skills by Transferring Experience. InProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023

work page 2023
[27]

Domain randomization for transferring deep neural networks from simulation to the real world

Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23–30. IEEE, 2017

work page 2017
[28]

More: Mixture of residual experts for humanoid lifelike gaits learning on complex terrains, 2025

Dewei Wang, Xinmiao Wang, Xinzhe Liu, Jiyuan Shi, Yingnan Zhao, Chenjia Bai, and Xuelong Li. More: Mixture of residual experts for humanoid lifelike gaits learning on complex terrains, 2025

work page 2025
[29]

Multi-expert learning of adaptive legged locomotion.Science Robotics, 5(49):eabb2174, 2020

Chuanyu Yang, Kai Yuan, Qiuguo Zhu, Wanming Yu, and Zhibin Li. Multi-expert learning of adaptive legged locomotion.Science Robotics, 5(49):eabb2174, 2020

work page 2020
[30]

Neural volumetric mem- ory for visual locomotion control

Ruihan Yang, Ge Yang, and Xiaolong Wang. Neural volumetric mem- ory for visual locomotion control. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1430– 1440, 2023

work page 2023
[31]

Learning vision-guided quadrupedal locomotion end- to-end with cross-modal transformers

Ruihan Yang, Minghao Zhang, Nicklas Hansen, Huazhe Xu, and Xiaolong Wang. Learning vision-guided quadrupedal locomotion end- to-end with cross-modal transformers. InInternational Conference on Learning Representations, 2022

work page 2022
[32]

Visual- locomotion: Learning to walk on complex terrains with vision

Wenhao Yu, Deepali Jain, Alejandro Escontrela, Atil Iscen, Peng Xu, Erwin Coumans, Sehoon Ha, Jie Tan, and Tingnan Zhang. Visual- locomotion: Learning to walk on complex terrains with vision. In5th Annual Conference on Robot Learning, 2021

work page 2021
[33]

Learn- ing agile locomotion on risky terrains

Chong Zhang, Nikita Rudin, David Hoeller, and Marco Hutter. Learn- ing agile locomotion on risky terrains. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11864– 11871. IEEE, 2024

work page 2024