Pith. sign in

REVIEW 3 major objections 6 minor 16 references

PolygMap: A Perceptive Locomotion Framework for Humanoid Robot Stair Climbing

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A humanoid robot can climb unknown staircases online by planning each footstep on a real-time polygonal map built from fused LiDAR, RGB-D, and IMU data.

desk verdict A real stair-climbing demo but Eq. 31, taken literally, zeroes foot height; worth refereeing for the hardware results, but the planning equations and the quantitative claims need hard fixes. read the letter →

arxiv 2510.12346 v2 pith:2SJETRSU submitted 2025-10-14 cs.RO

Bingquan Li , Ning Wang , Zhicheng He , Yucong Wu , Tianwei Zhang This is my paper · ORCID
classification cs.RO
keywords humanoidlocomotionstairclimbingperceptivepolygonmappingfootstepplanningmulti-sensorfusionLiDAR-inertialodometryRGB-Dperception
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

The paper proposes PolygMap, a perception–planning framework that lets a full-size humanoid robot climb multi-step staircases it has never seen before. The key move is to turn raw depth and LiDAR measurements into a compact polygon map of stair tread planes in real time, then erode those polygons to obtain safe foothold candidates and plan whole-body stepping motions at 20–30 Hz on an onboard computer. The authors validate the framework in Gazebo simulation and in real indoor and outdoor stair-climbing experiments, reporting footstep errors of 11–33 mm and successful climbs of 4–6 consecutive steps. They also report that the faster single-step gait has a nonzero failure rate outdoors, attributed to state-estimation inaccuracy, lighting, and actuator issues under that gait. If correct, the work demonstrates that online foothold planning from a compact polygonal representation can support sustained stair climbing without pre-mapped geometry.

What carries the argument

The load-bearing object is the eroded polygonal foothold map: a set of convex-hull polygons fitted to detected stair planes, rasterized into a dense point cloud, filtered to the robot's reachable range and foot-height threshold, and eroded per height layer to suppress edge noise. This representation converts sparse, noisy depth measurements into stable candidate foothold positions near the robot, and it is what couples perception to footstep planning. The supporting machinery is the state estimator: a linear Kalman filter plus complementary filtering that fuses LiDAR-inertial odometry with forward kinematics so that depth polygons are anchored to a low-drift body pose.

What would settle it

Run the same pipeline on a staircase with matte black or otherwise absorptive treads in outdoor sunlight: if the polygonal map misses treads or produces false planes, foothold candidates disappear and the robot either stops or clips a stair edge, which would falsify the claim of robust unknown-stair climbing. A less extreme check: measure planned-versus-executed footstep error on a long single-step outdoor climb and see whether it exceeds the tread-width minus foot-length safety margin (about 2 cm on the reported 28 cm tread and 26 cm foot).

Watch

Extended reading notes

Core claim

The central claim is that a polygonal staircase plane semantic map, rebuilt at 20–30 Hz from depth images and fused LiDAR-inertial/proprioceptive odometry, can serve as the sole geometric basis for humanoid stair-climbing footstep planning. The pipeline extracts plane polygons from the depth stream using anisotropic diffusion, Sobel-based normals, Canny contours, and RANSAC fitting, anchors them to the world with a low-drift fused pose, rasterizes each polygon into a dense point cloud, erodes the points by height layer to remove edge noise, and selects the closest reachable foothold that satisfies the step-height threshold. The authors report that this closed loop climbs 4–6 consecutive real

Load-bearing premise

The whole loop depends on the downward-facing depth camera producing accurate polygons of stair treads while the robot is walking; the paper itself notes the sensor degrades on highly absorptive surfaces and blames some outdoor failures on lighting and state-estimation inaccuracy.

Editorial extensions

If this is right

  • A full-size humanoid can climb multi-step, previously unseen staircases without a pre-built map, provided the onboard depth stream yields reliable tread polygons while walking.
  • The 20–30 Hz planning rate means foothold decisions can be recomputed within a single step cycle, allowing gait adjustment to the next detected tread.
  • The reported 11–33 mm planned-versus-executed footstep errors are small relative to the 28-cm tread width, implying a comfortable safety margin for the double-step gait.
  • The faster single-step gait trades accuracy for speed, with larger errors approaching the safety margin, so its use on long outdoor stairs requires better state estimation or gait timing.

Reading between the lines

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

  • The admitted outdoor failures point to the depth sensor's known weakness on dark, absorptive tread surfaces as the likeliest single-point failure; a depth modality robust to matte black surfaces would directly extend the method to more real stairwells.
  • The same polygon-erosion pipeline could generalize beyond stairs to gap crossing or stepping stones, since it already reduces arbitrary planar regions to ranked foothold candidates with a height threshold.
  • A testable extension would be to add predictive foothold selection during the swing phase of the single-step gait, using the fused state estimate to plan the next step before the body has fully settled, which could shrink the 33.4 mm error.
  • The paper leaves implicit that the 20–30 Hz polygon map rate is measured on an NVIDIA Orin; porting to lighter embedded hardware would clarify the minimum compute budget for perceptive stair climbing.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents PolygMap, a perception and locomotion framework for humanoid stair climbing. It fuses Point-LIO LiDAR-inertial odometry with joint kinematics through a linear Kalman filter and complementary filtering, builds a polygonal stair map from a downward-facing RGB-D camera via anisotropic diffusion, Sobel normals, Canny/RANSAC plane fitting, converts polygons into dense point clouds with erosion-based foothold candidates, and generates sinusoidal foot trajectories. Validation is performed in Gazebo and on a KUA VO humanoid, with indoor/outdoor stair climbs using double-step and single-step gaits. The paper reports planning rates of 20–30 Hz, footstep errors of 11.4–33.4 mm in real trials, and one documented outdoor toe-collision failure.

Significance. If the described methods work as claimed, the paper makes a useful systems contribution: full-size humanoid robots can climb unknown multi-step staircases online using onboard perception, without pre-mapped models, at planning rates high enough for real-time control. The inclusion of quantitative footstep errors, simulation trials, and a candid description of a failure mode is a strength. However, the support is limited: only five real trials are reported, no failure-rate or variance statistics are given, and the planning formulation in Sec. III-C appears to have a missing coupling between the detected foothold height and the generated foot trajectory. The central idea is plausible, but the manuscript as written does not yet provide a reproducible description.

major comments (3)
  1. [III-C, Eqs. (31)–(32)] The footstep planner as written cannot place feet on stair treads. With p_t = [x_t, y_t, z_t, φ_t]^T and b_f = [0, ±y_b, −z_t]^T, Eq. (31) gives p_{f,z} = z_t − z_t = 0 for every footstep. The Z-trajectory in Eq. (32) introduces a 'reference landing height' z_0, but z_0 is never defined as the height of the selected foothold from Sec. III-B, and no equation connects the output p* of Eq. (26) to x_end, y_end, or z_0 in Eqs. (32)–(33). Thus the described pipeline is not sufficient to reproduce the reported climbing of 13 cm steps; either Eq. (31) is a typo or the foothold height from Sec. III-B is never used. This coupling must be specified explicitly.
  2. [III-B to III-C] The foothold generation stage outputs p* (Eq. 26), but the footstep planning stage takes only the torso pose B_t as input (Eq. 29). Neither x_end/y_end in Eq. (33) nor z_0 in Eq. (32) is defined in terms of p* or the polygon map. Consequently, even if Eq. (31)'s z-component were corrected, the text would still not specify how the detected stair geometry determines the target foot placement. The authors should give the explicit mapping from the selected foothold polygon to the footstep target used in Eqs. (31)–(33).
  3. [IV-C, Table II and Fig. 8] The real-world validation consists of five runs (two indoor DS, three SS, one of which failed). No error bars, confidence intervals, or failure-rate statistics are reported, although the text acknowledges a 'failure rate' for outdoor SS. Given the robot's foot length (26 cm) and the stair tread (28 cm), a 33.4 mm maximum error can be comparable to or greater than the available 20 mm longitudinal margin; the connection between the reported errors and the observed toe collision should be analyzed quantitatively. As it stands, the claim of robustness is stronger than the data support.
minor comments (6)
  1. [Abstract] 'footstep planar' should be 'footstep planner'.
  2. [Fig. 3] The figure contains the typo 'Kallman' for 'Kalman'.
  3. [Fig. 4 caption] 'Ontology-Based' appears to be a typo for 'Odometry-Based'.
  4. [III-B, Eq. (14)] Setting the height of each polygon vertex to zero before computing the 2D convex hull is deliberate, but it may confuse readers; clarify that the original heights are restored in Eq. (18).
  5. [III-B, last paragraph] The text says the optimal foothold points provide 'both spatial position and orientation,' but Eq. (26) returns only a position. Specify how the foot yaw is determined.
  6. [References] Reference [15] contains a URL with a space ('polygon mapping') and should be formatted correctly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is self-contained and external dependencies are prior published systems.

full rationale

The claimed derivation chain does not reduce to its inputs. State estimation combines forward kinematics with Point-LIO odometry using fixed complementary/Kalman filtering coefficients, not fitted to the reported stair-climbing success. Perception converts depth images to surface normals, RANSAC plane fits, and eroded foothold candidates; foothold selection is an explicit geometric pipeline from sensor data, not a parameter fitted to the final result and then called a prediction. The footstep planner consumes the foothold output and generates trajectories, but no equation defines the foothold output in terms of the trajectory or vice versa. The only external dependencies are cited prior systems (Point-LIO, a GitHub polygon-mapping repository), and neither is authored by the present paper's authors, so no load-bearing self-citation chain is present. The apparent inconsistency in Sec. III-C, where Eq. 31 gives p_f,z = z_t + (−z_t) = 0 while Eq. 32 introduces a separate landing height z_0, is a reproducibility/correctness concern rather than a circularity: the foot height is not being derived from itself, it is simply disconnected from the foothold height selected in Sec. III-B. Accordingly, no circular step can be quoted, and the honest finding is no significant circularity.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central result depends less on novel mathematics than on hand-chosen perception thresholds and standard assumptions about stair geometry and sensor accuracy. None of the thresholds are given numerical values in the text, so the reader cannot reproduce the behavior without additional information.

free parameters (8)
  • Complementary time constant τ (or α) = not specified
    Controls blend ratio between proprioceptive and LIO pose in Eqs. (6)-(9); hand-tuned, directly shapes state-estimation smoothness and lag.
  • Anisotropic diffusion coefficient λ and iteration count = not specified
    Appears in Eq. (13); determines degree of depth-image denoising vs edge preservation, chosen without reported values.
  • Grid resolution g_res = not specified
    Used to rasterize convex hull polygons into dense points (Eqs. 16-21); sets foothold candidate density and horizontal accuracy.
  • Range gate g_range = not specified
    Eq. (20) restricts foothold search to a rectangle around the robot; chosen by hand and affects reachability.
  • Foot height threshold g_z = not specified
    Eq. (23) removes points above foot height plus threshold; defines the maximum climbable step height.
  • Layer height h_layer and erosion count N_erosion = not specified
    Eqs. (24)-(25) discretize point clouds into height layers and erode each layer; noise/edge filtering parameters, values not given.
  • Foothold height margin Δ_foot = not specified
    Eqs. (26)/(28) select footholds above current sole height; central to whether a step is accepted as climbable.
  • Swing timing t_lift, t_land, maximum step height z_max = not specified
    Eq. (32) shapes the sinusoidal foot-lift trajectory; hand-specified gait parameters.
assumptions (6)
  • domain assumption Stair treads are planar and roughly horizontal; average vertex height represents tread height.
    Eqs. (14), (18) project polygons to z=0 and assign mean z to all rasterized points; if treads are curved or sloped, foothold heights are biased.
  • domain assumption Surface normals from cross-products of neighboring depth pixels (Eq. 12) are reliable after anisotropic diffusion.
    Under motion blur, rolling shutter, or low texture, this assumption can break; acknowledged in Sec. III and through outdoor failures.
  • domain assumption Point-LIO provides drift-free absolute pose and the LiDAR-to-base extrinsic is exact.
    Eq. (5) transforms LIO pose to robot CoM; any calibration error or odometry drift propagates into the polygon map.
  • domain assumption Linear Kalman filter with constant-velocity model (Eq. 2) adequately describes humanoid torso and foot-contact motion.
    Used in Sec. III-A; impacts fused pose smoothness and foothold localization.
  • domain assumption Convex hull and layered erosion of foot-height layers retain all safe footholds.
    Sec. III-B; on concave stair edges or small irregular treads, convex hull can fill unsafe regions that erosion may not remove.
  • domain assumption Rotating-rectangle foot model (Eq. 34) is sufficient to detect foot overlap and collisions.
    Ignores toe/heel shape and z-collisions; used for trajectory feasibility in Sec. III-C.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PolygMap: A Perceptive Locomotion Framework for Humanoid Robot Stair Climbing." pith.science (2026). https://pith.science/paper/2SJETRSU

@misc{pith2026251012346,
  author       = {Pith},
  title        = {Pith review of: PolygMap: A Perceptive Locomotion Framework for Humanoid Robot Stair Climbing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SJETRSU}},
  note         = {Machine review of arXiv:2510.12346}
}
read the original abstract

Recently, biped robot walking technology has been significantly developed, mainly in the context of a bland walking scheme. To emulate human walking, robots need to step on the positions they see in unknown spaces accurately. In this paper, we present PolyMap, a perception-based locomotion planning framework for humanoid robots to climb stairs. Our core idea is to build a real-time polygonal staircase plane semantic map, followed by a footstep planar using these polygonal plane segments. These plane segmentation and visual odometry are done by multi-sensor fusion(LiDAR, RGB-D camera and IMUs). The proposed framework is deployed on a NVIDIA Orin, which performs 20-30 Hz whole-body motion planning output. Both indoor and outdoor real-scene experiments indicate that our method is efficient and robust for humanoid robot stair climbing.

Figures

Figures reproduced from arXiv: 2510.12346 by the authors.

Figure 1
Figure 1. Perceptive Locomotion for Humanoid Stair Climbing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The system integrates joint recorders, depth sensing and LIO estimator. Robot pose is obtained via fusing forward [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Polygmap-based footstep motion planning logic [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Simulation experiment of the proposed system. The [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Real-world stair climbing experiment. The figure presents the experimental setup and key motion snapshots during [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Experimental results of real-world stair climbing. The left plot shows the 3D trajectory of the robot base and footsteps, [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Two outdoor experiment scenes. These snapshots [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references

  1. [1]

    Multicontact locomotion of legged robots,

    J. Carpentier and N. Mansard, “Multicontact locomotion of legged robots,”IEEE Transactions on Robotics, vol. 34, no. 6, pp. 1441– 1460, 2018. 1

    2018

  2. [2]

    Stairs detection with odometry-aided traversal from a wearable rgb-d camera,

    A. Perez-Yus, D. Gutierrez-Gomez, G. Lopez-Nicolas, and J. Guerrero, “Stairs detection with odometry-aided traversal from a wearable rgb-d camera,”Computer Vision and Image Understanding, vol. 154, pp. 192–205, 2017. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S1077314216300315 1

    2017

  3. [3]

    Fast cylinder and plane extraction from depth cameras for visual odometry,

    P. F. Proenc ¸a and Y . Gao, “Fast cylinder and plane extraction from depth cameras for visual odometry,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 6813–

    2018

  4. [4]

    Plane segmenta- tion using depth-dependent flood fill,

    A. Roychoudhury, M. Missura, and M. Bennewitz, “Plane segmenta- tion using depth-dependent flood fill,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 2210–

    2021

  5. [5]

    Gpu-accelerated rapid planar region extraction for dynamic be- haviors on legged robots,

    B. Mishra, D. Calvert, S. Bertrand, S. McCrory, R. Griffin, and H. E. Sevil, “Gpu-accelerated rapid planar region extraction for dynamic be- haviors on legged robots,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 8493–8499. 2

    2021

  6. [6]

    Polytopic planar region characterization of rough terrains for legged locomotion,

    Z. Xu, H. Zhu, H. Chen, and W. Zhang, “Polytopic planar region characterization of rough terrains for legged locomotion,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 8682–8689. 2

    2022

  7. [7]

    Stair climbing stabilization of the hrp-4 humanoid robot using whole-body admittance control,

    S. Caron, A. Kheddar, and O. Tempier, “Stair climbing stabilization of the hrp-4 humanoid robot using whole-body admittance control,” in 2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 277–283. 2

    2019

  8. [8]

    A ground-stair walking strategy of the assistive device based on the rgb-d camera,

    S.-H. Yu, B.-R. Yang, H.-H. Lee, and E. Tanaka, “A ground-stair walking strategy of the assistive device based on the rgb-d camera,” in2021 IEEE/SICE International Symposium on System Integration (SII), 2021, pp. 341–346. 2

    2021

Show all 16 references
  1. [9]

    3d polygonal mapping for humanoid robot navigation,

    A. Roychoudhury, M. Missura, and M. Bennewitz, “3d polygonal mapping for humanoid robot navigation,” in2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), 2022, pp. 171–177. 2

    2022

  2. [10]

    Drift-free humanoid state estimation fusing kinematic, inertial and lidar sensing,

    M. F. Fallon, M. Antone, N. Roy, and S. Teller, “Drift-free humanoid state estimation fusing kinematic, inertial and lidar sensing,” in2014 IEEE-RAS International Conference on Humanoid Robots. IEEE, 2014, pp. 112–119. 2

    2014

  3. [11]

    Elevation mapping for locomotion and navigation using gpu,

    T. Miki, L. Wellhausen, R. Grandia, F. Jenelten, T. Homberger, and M. Hutter, “Elevation mapping for locomotion and navigation using gpu,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 2273–2280. 2

    2022

  4. [12]

    Efficient terrain map using planar regions for footstep planning on humanoid robots,

    B. Mishra, D. Calvert, S. Bertrand, J. Pratt, H. E. Sevil, and R. Griffin, “Efficient terrain map using planar regions for footstep planning on humanoid robots,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 8044–8050. 2

    2024

  5. [13]

    Perceptive locomotion through nonlinear model-predictive control,

    R. Grandia, F. Jenelten, S. Yang, F. Farshidian, and M. Hutter, “Perceptive locomotion through nonlinear model-predictive control,” IEEE Transactions on Robotics, vol. 39, no. 5, pp. 3402–3421, 2023. 2

    2023

  6. [14]

    Attention-based map encoding for learning generalized legged loco- motion,

    J. He, C. Zhang, F. Jenelten, R. Grandia, M. B ¨acher, and M. Hutter, “Attention-based map encoding for learning generalized legged loco- motion,”Science Robotics, vol. 10, no. 105, p. eadv3604, 2025. 2

    2025

  7. [15]

    T. Bin. (2024) Rgb-d polygon mapping. [Online]. Available: https://github.com/BTFrontier/polygon mapping 3

    2024

  8. [16]

    Point-lio: Robust high-bandwidth light detection and ranging inertial odometry,

    D. He, W. Xu, N. Chen, F. Kong, C. Yuan, and F. Zhang, “Point-lio: Robust high-bandwidth light detection and ranging inertial odometry,” Advanced Intelligent Systems, vol. 5, no. 7, p. 2200459, 2023. 3

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.