pith. sign in

arxiv: 2605.20484 · v1 · pith:HMYBGHWXnew · submitted 2026-05-19 · 💻 cs.RO

Enhancing Graph-Based SLAM in GNSS-Denied environments by leveraging leg odometry

L\'eon Perruchot-Triboulet , Luc Jaulin , Kai Xiao This is my paper

Pith reviewed 2026-05-21 06:50 UTC · model grok-4.3

classification 💻 cs.RO
keywords Graph-based SLAMLeg odometryFactor graphGNSS-denied navigationQuadruped robotLIO-SAMElevation driftProprioceptive sensing
0
0 comments X

The pith

Augmenting LIO-SAM with leg odometry reduces elevation drift from over 30 m to under 30 cm on a quadruped.

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

The paper develops a factor-graph method that adds a separate lane using leg odometry to supply vertical position data when GNSS is unavailable and LiDAR alone drifts in elevation. This lane connects to the main LiDAR-inertial pipeline through an identity relative-pose constraint that applies noise selectively so only vertical components are anchored. Experiments on a quadruped robot covering two outdoor loops longer than one kilometer show the map stays accurate and the optimizer converges in a scene where the unaugmented system fails. A reader would care because the method re-uses motion data the robot already computes for walking, avoiding extra sensors or heavy computation.

Core claim

A factor graph architecture augments the LIO-SAM framework with a parallel kinematic lane driven by proprioceptive leg odometry; the two lanes couple through an identity relative-pose constraint equipped with a selective noise model, which reduces elevation drift from over 30 m to under 30 cm and produces convergence on kilometer-scale outdoor loops where the baseline pipeline fails entirely.

What carries the argument

Parallel kinematic lane from proprioceptive leg odometry coupled to the LiDAR-inertial lane by an identity relative-pose constraint with selective noise model.

If this is right

  • The augmented graph converges in geometrically sparse or repetitive scenes where standard LIO-SAM fails.
  • Vertical position remains accurate to under 30 cm across trajectories longer than one kilometer.
  • Proprioceptive data already computed for gait control can serve as a lightweight vertical anchor without extra hardware.
  • Selective noise on the identity constraint limits its influence to the vertical direction.

Where Pith is reading between the lines

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

  • The same vertical-anchor idea could be tested on other legged platforms or fused with additional proprioceptive signals such as joint torque estimates.
  • Early integration of internal motion estimates might reduce reliance on vertical loop closures in future graph-SLAM pipelines.
  • Applying the selective-noise identity constraint to other low-drift proprioceptive sources would test how general the vertical-stabilization effect is.

Load-bearing premise

Leg odometry supplies sufficiently accurate vertical information that fuses as an identity relative-pose constraint with selective noise without injecting new systematic bias into the graph solution.

What would settle it

Re-running the same outdoor kilometer loops with the leg-odometry lane disabled and measuring whether elevation drift again exceeds 30 m or the optimizer fails to converge would falsify the performance gain.

Figures

Figures reproduced from arXiv: 2605.20484 by Kai Xiao, L\'eon Perruchot-Triboulet, Luc Jaulin.

Figure 1
Figure 1. Figure 1: Map generated with the proposed modified SLAM framework on both experimental datasets. Each image is [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SLAM pipeline crash with visible false elevation [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hybrid factor graph with a parallel lane leg [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
read the original abstract

Autonomous navigation in GNSS-denied environments remains a core challenge for legged robots, where exteroceptive sensors such as LiDAR are prone to elevation drift in geometrically sparse or repetitive scenes. We present a factor graph architecture that augments the LIO-SAM framework with a parallel kinematic lane driven by proprioceptive leg odometry, coupled to the main LiDAR-inertial lane via an identity relative pose constraint with a selective noise model. Applied to a Linxai D50 quadruped platform across two outdoor loops totaling over one kilometer, our approach reduces elevation drift from over 30m to under 30cm and enables convergence in a scene where the baseline pipeline fails entirely. These results suggest that proprioceptive data, already computed onboard for gait control, constitutes a lightweight and effective vertical anchor for SLAM in GNSS-denied settings.

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

2 major / 2 minor

Summary. The manuscript augments the LIO-SAM factor-graph SLAM pipeline with a parallel kinematic lane driven by proprioceptive leg odometry from a quadruped. The two lanes are coupled by an identity relative-pose constraint equipped with a selective noise model (low vertical variance). Experiments on a Linxai D50 platform over two outdoor loops totaling more than one kilometer report reduction of elevation drift from >30 m to <30 cm and successful convergence where the baseline pipeline diverges.

Significance. If the central experimental claim holds under independent verification, the work shows that leg odometry already computed for gait control can provide a lightweight vertical anchor for graph-based SLAM in GNSS-denied terrain. This is a practical contribution for legged-robot navigation without extra sensors.

major comments (2)
  1. [factor-graph architecture description] Factor-graph architecture description: the identity relative-pose constraint with selective noise model assumes leg-odometry vertical estimates remain unbiased relative to the inertial frame over kilometer-scale trajectories. No independent quantification of leg-odometry vertical error (e.g., against external ground truth or IMU) or ablation removing the selective noise model is reported, so it is unclear whether the observed <30 cm drift is a genuine correction or an artifact of the added constraint.
  2. [experimental results] Experimental evaluation: the headline result (elevation drift reduced from >30 m to <30 cm over >1 km) is presented without error bars, exact loop-closure selection criteria, or a full protocol for data exclusion. This leaves the quantitative claim only partially verifiable and weakens the assertion that the method enables convergence where the baseline fails entirely.
minor comments (2)
  1. [abstract] Abstract: the total distance and platform name are already stated; consider also reporting the number of trials or the specific terrain types for immediate context.
  2. [throughout] Notation: ensure the terms 'kinematic lane' and 'parallel lane' are used consistently when referring to the leg-odometry factor-graph component.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of the factor-graph design and experimental reporting that we address below. We have revised the manuscript accordingly to improve clarity and verifiability.

read point-by-point responses

  1. Referee: Factor-graph architecture description: the identity relative-pose constraint with selective noise model assumes leg-odometry vertical estimates remain unbiased relative to the inertial frame over kilometer-scale trajectories. No independent quantification of leg-odometry vertical error (e.g., against external ground truth or IMU) or ablation removing the selective noise model is reported, so it is unclear whether the observed <30 cm drift is a genuine correction or an artifact of the added constraint.

    Authors: The selective noise model is grounded in the kinematic properties of leg odometry on a quadruped, where vertical displacement is computed directly from leg joint angles and foot contact events, providing a terrain-relative anchor that is less prone to bias than horizontal components affected by slippage. We acknowledge the value of an independent quantification against external sensors. While such isolated leg-odometry ground-truth data was not collected in the original experiments, we have added an ablation study in the revised manuscript that disables the selective noise model and demonstrates increased elevation drift, confirming the constraint's contribution. We have also expanded the architecture description to better justify the modeling assumptions. revision: partial

  2. Referee: Experimental evaluation: the headline result (elevation drift reduced from >30 m to <30 cm over >1 km) is presented without error bars, exact loop-closure selection criteria, or a full protocol for data exclusion. This leaves the quantitative claim only partially verifiable and weakens the assertion that the method enables convergence where the baseline fails entirely.

    Authors: We have updated the experimental section to specify the exact loop-closure detection parameters and to provide a complete description of the data processing pipeline, confirming that no additional data exclusion was applied beyond the standard LIO-SAM configuration. The two reported loops represent the full trajectories collected. We have added a note on result consistency across the loops to address variability; given that the evaluation uses complete real-world recordings rather than repeated trials with randomized elements, formal error bars are not applicable, but the qualitative convergence improvement remains evident. revision: yes

standing simulated objections not resolved
  • Independent quantification of leg-odometry vertical error against external ground truth or IMU, as this would require additional sensor data and experiments not performed in the original study.

Circularity Check

0 steps flagged

No circularity; results are empirical outcomes from external robot experiments

full rationale

The paper augments LIO-SAM with a parallel leg-odometry lane connected by an identity relative-pose factor and selective noise model. The headline performance numbers (elevation drift reduced from >30 m to <30 cm over 1 km loops on a Linxai D50 platform) are presented as direct measurements from physical trials, not as outputs of any closed-form derivation or fitted parameter that is then relabeled as a prediction. No equations, self-citations, or ansatzes are shown that would reduce the reported improvement to the input constraints by construction. The architecture is therefore self-contained against the external benchmark of real-world robot data.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility; the main unstated premise is that leg odometry remains unbiased in the vertical axis under the tested outdoor conditions.

axioms (1)
  • domain assumption Leg odometry provides sufficiently accurate vertical pose information for use as an identity constraint.
    Invoked when coupling the kinematic lane to the LiDAR-inertial lane.

pith-pipeline@v0.9.0 · 5679 in / 1159 out tokens · 45087 ms · 2026-05-21T06:50:44.927801+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

10 extracted references · 10 canonical work pages

  1. [1]

    LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping , year =

    Shan, Tixiao and Englot, Brendan and Meyers, Drew and Wang, Wei and Ratti, Carlo and Rus, Daniela , booktitle =. LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping , year =

    work page

  2. [2]

    Robotics , volume=

    State estimation for legged robots-consistent fusion of leg kinematics and IMU , author=. Robotics , volume=. 2013 , publisher=

    work page 2013

  3. [3]

    Fallon , title =

    David Wisth and Marco Camurri and Maurice F. Fallon , title =. CoRR , volume =. 2021 , url =. 2107.07243 , timestamp =

    work page arXiv 2021

  4. [4]

    2022 , publisher =

    Frank Dellaert and GTSAM Contributors , title =. 2022 , publisher =. doi:10.5281/zenodo.5794541 , url =

    work page doi:10.5281/zenodo.5794541 2022

  5. [5]

    iSAM2: Incremental smoothing and mapping with fluid relinearization and incremental variable reordering , year=

    Kaess, Michael and Johannsson, Hordur and Roberts, Richard and Ila, Viorela and Leonard, John and Dellaert, Frank , booktitle=. iSAM2: Incremental smoothing and mapping with fluid relinearization and incremental variable reordering , year=

    work page

  6. [6]

    2019 , booktitle =

    Incremental visual-inertial 3d mesh generation with structural regularities , author =. 2019 , booktitle =

    work page 2019

  7. [7]

    LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain , year=

    Shan, Tixiao and Englot, Brendan , booktitle=. LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain , year=

    work page

  8. [8]

    Tightly-Coupled LiDAR-IMU-Wheel Odometry with an Online Neural Kinematic Model Learning via Factor Graph Optimization , doi =

    Okawara, Taku and Koide, Kenji and Oishi, Shuji and Yokozuka, Masashi and Banno, Atsuhiko and Uno, Kentaro and Yoshid, Kazuya , year =. Tightly-Coupled LiDAR-IMU-Wheel Odometry with an Online Neural Kinematic Model Learning via Factor Graph Optimization , doi =

    work page

  9. [9]

    2021 , eprint=

    FAST-LIO2: Fast Direct LiDAR-inertial Odometry , author=. 2021 , eprint=

    work page 2021

  10. [10]

    Li, Kailai and Li, Meng and Hanebeck, Uwe D. , year=. Towards High-Performance Solid-State-LiDAR-Inertial Odometry and Mapping , volume=. IEEE Robotics and Automation Letters , publisher=. doi:10.1109/lra.2021.3070251 , number=

    work page doi:10.1109/lra.2021.3070251 2021