pith. sign in

arxiv: 2605.00261 · v1 · submitted 2026-04-30 · 💻 cs.RO

Task-Conditioned Uncertainty Costmaps for Legged Locomotion

Kartikeya Singh , Christo Aluckal , Romeo Orsolino , Karthik Dantu This is my paper

Pith reviewed 2026-05-09 19:41 UTC · model grok-4.3

classification 💻 cs.RO
keywords legged locomotionepistemic uncertaintyfoothold predictioncostmapsout-of-distribution detectionmotion planningterrain perception
0
0 comments X

The pith

Modeling epistemic uncertainty in foothold predictions lets legged robots flag unfamiliar terrain and plan more reliably.

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

The paper establishes that a foothold prediction model can be trained to output not only contact locations but also the epistemic uncertainty in those predictions when conditioned on terrain height scans and the commanded task. This uncertainty distinguishes regions the model has seen during training from those it has not, even when the terrain looks geometrically similar. By folding the uncertainty values into a unified costmap, the planner avoids or penalizes high-uncertainty footholds, which the authors show reduces feasibility errors in both simulation and hardware experiments. A single model trained on limited distributions can therefore operate across a wider range of unstructured environments without retraining.

Core claim

Epistemic uncertainty in predicted footholds, conditioned on terrain observations and commanded motion, distinguishes in-distribution from out-of-distribution operating regimes. This uncertainty is incorporated into costmap generation, yielding improved OOD detection and up to 37 percent lower simulation feasibility error compared with geometry-only baselines, together with more reliable planning behavior on real hardware.

What carries the argument

Task-conditioned epistemic uncertainty estimator for foothold prediction, whose scalar outputs are rasterized into costmaps that augment standard geometric costs for motion planning.

If this is right

  • A single trained model expresses uncertainty caused by missing coverage rather than requiring separate OOD detectors.
  • Uncertainty-aware costmaps improve feasibility error by up to 37 percent in simulation across in-distribution and OOD terrains.
  • Real-world planning behavior becomes more reliable than geometry-only baselines when uncertainty is included.
  • The same uncertainty signal supports both OOD detection and path selection within one framework.

Where Pith is reading between the lines

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

  • The approach could be combined with online model updates that reduce uncertainty as new terrain data arrives.
  • High-uncertainty regions might trigger active exploration behaviors to collect training coverage before committing to a path.
  • Similar conditioning of uncertainty on task variables may apply to other contact-rich planning problems such as manipulation or aerial landing.

Load-bearing premise

The learned uncertainty values reliably indicate gaps in the training distribution and can be added to costmaps without introducing excessive conservatism or new failure modes.

What would settle it

A controlled test on terrain patches deliberately excluded from training data where uncertainty remains low yet planning still produces infeasible contacts, or where uncertainty is high yet the robot succeeds without incident.

Figures

Figures reproduced from arXiv: 2605.00261 by Christo Aluckal, Karthik Dantu, Kartikeya Singh, Romeo Orsolino.

Figure 1
Figure 1. Figure 1: Overview of the training pipeline Input Representation: At each timestep, the terrain geometry is rep￾resented as a 6 × 17 grid of ele￾vation values over a 0.6 m × 1.6 m frontal region, sampled at 0.1 m res￾olution, yielding 102 height measure￾ments (zi,j ). Grid Pooling: The frontal height grid is processed via non-overlapping average pooling into a compact ter￾rain descriptor fgrid ∈ R 12, capturing coar… view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE of terrain input distribu￾tion shift: training (red) vs. OOD test (blue). Left: ID features. Middle: OOD terrain, ID velocity. Right: OOD ter￾rain, OOD velocities [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Uncertainty–error correlation: terrain-only (left) vs. Proposed (right). We train a single model in Isaac￾Sim on flat terrain with constant velocities (ID set). Conditions in￾ducing distribution shift constitute OOD; elevated uncertainty under OOD is enforced via Lcal (Equa￾tion (7)). Baselines: (i) Terrain￾only uncertainty (Roughness): height￾scan-variance OOD detection treat￾ing high-variance regions as … view at source ↗
Figure 4
Figure 4. Figure 4: Terrain-only vs Proposed uncer￾tainty. Darker shade = uncertainty spike. We assess whether the proposed epistemic uncertainty reliably distin￾guishes ID from OOD terrain re￾gions, and whether detected OOD re￾gions correspond to genuinely harder footholds. At each timestep t, the pro￾posed method computes the scalar epistemic summary ¯st (from eq. (3)) and the terrain-only baseline com￾putes height-scan var… view at source ↗
Figure 5
Figure 5. Figure 5: KDE of goal-progress across 20 different MPPI planning runs [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Progressive effect of uncertainty weighting on planning. From (1) to (4), increasing uncertainty cost shifts the planner from aggressive, risk-prone trajectories to conservative paths that avoid uncertain terrain. We evaluate the feasibility of the path traversed by the robot as ex￾plained in Section 5. In this experiment, we compare our proposed path with two baselines: (i) obstacle-only, which uses a fix… view at source ↗
Figure 7
Figure 7. Figure 7: Planning results using NAV2 on Unitree Go1 robot. The purple patch marks the anomaly induced by the OOD velocity and roughness, created by plac￾ing a 5cm board along the path [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
read the original abstract

Legged robots maintain dynamic feasibility through multicontact interactions with terrain. Learned foothold prediction can provide feasibility-aware costs for motion planning and path selection, but accurately predicting future contacts from perceptual inputs such as height scans remains challenging on highly unstructured terrain, even with a repetitive gait cycle. In this work, we show that modeling epistemic uncertainty in predicted footholds, conditioned on terrain observations and commanded motion, distinguishes in-distribution from out-of-distribution operating regimes in simulation and real-world settings. This allows a single learned model, trained on limited data distributions, to express uncertainty caused by missing training coverage. We use this learned uncertainty to detect OOD regions and incorporate them into a unified costmap-generation framework for uncertainty-aware path planning. Using these uncertainty-aware costmaps, we evaluate feasibility error across in-distribution and OOD terrains in simulation and real-world settings. The results show improved OOD detection, up to a 37% reduction in simulation feasibility error, and more reliable planning behavior than geometry-only baselines.

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 / 1 minor

Summary. The paper claims that modeling epistemic uncertainty in predicted footholds—conditioned on terrain observations and commanded motion—allows a single learned model to distinguish in-distribution from out-of-distribution regimes for legged robots. This uncertainty is incorporated into a unified costmap framework for uncertainty-aware path planning, yielding improved OOD detection, up to 37% reduction in simulation feasibility error, and more reliable planning than geometry-only baselines in both simulation and real-world experiments.

Significance. If validated, the work offers a practical way to deploy limited-data learned foothold predictors safely on unstructured terrain by explicitly signaling missing coverage via epistemic uncertainty. The task-conditioned formulation and integration into costmaps for planning represent a useful advance for legged locomotion, with the dual sim/real evaluation adding weight to the feasibility claims.

major comments (2)
  1. [Results] Results section: the central claim that epistemic uncertainty specifically flags missing training coverage (rather than in-distribution terrain difficulty) is load-bearing for interpreting the 37% feasibility-error reduction and OOD detection gains as evidence of true distribution awareness. No ablation or analysis is described that compares uncertainty values on complex but in-distribution terrains versus explicitly constructed OOD cases, leaving open the alternative that gains arise from added conservatism.
  2. [Methods] Methods section: the uncertainty estimation procedure (ensemble size, dropout rate, or other technique), model architecture, training data distribution, and exact definition of the feasibility error metric are not detailed enough to assess whether the reported gains are robust or reproducible; these elements are required to evaluate support for the OOD-signaling claim.
minor comments (1)
  1. [Abstract] Abstract: quantitative claims (37% error reduction, improved OOD detection) are stated without naming the precise metric, baseline methods, or number of trials, which reduces immediate clarity even though full details presumably appear later.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and have revised the manuscript to strengthen the presentation of our results and methods.

read point-by-point responses

  1. Referee: [Results] Results section: the central claim that epistemic uncertainty specifically flags missing training coverage (rather than in-distribution terrain difficulty) is load-bearing for interpreting the 37% feasibility-error reduction and OOD detection gains as evidence of true distribution awareness. No ablation or analysis is described that compares uncertainty values on complex but in-distribution terrains versus explicitly constructed OOD cases, leaving open the alternative that gains arise from added conservatism.

    Authors: We agree that an explicit comparison of uncertainty values on complex in-distribution terrains versus OOD cases is important to rule out the possibility that uncertainty simply reflects terrain difficulty. Our original experiments already include both complex ID terrains and constructed OOD examples (novel terrain geometries outside the training distribution), with quantitative gains reported on the latter. However, we did not present a dedicated side-by-side ablation of uncertainty magnitudes. In the revised manuscript we have added a new analysis and figure in the Results section that directly compares mean epistemic uncertainty across increasing terrain complexity within the training distribution against the same metric on OOD cases. The added results show that uncertainty remains comparatively low on complex but in-distribution terrain while rising markedly on OOD terrain, supporting the distribution-awareness interpretation. We have also clarified the construction of the OOD test sets in the text. revision: yes

  2. Referee: [Methods] Methods section: the uncertainty estimation procedure (ensemble size, dropout rate, or other technique), model architecture, training data distribution, and exact definition of the feasibility error metric are not detailed enough to assess whether the reported gains are robust or reproducible; these elements are required to evaluate support for the OOD-signaling claim.

    Authors: We concur that these details are necessary for reproducibility and for readers to evaluate the strength of the OOD-signaling results. In the revised Methods section we now provide: (i) the uncertainty estimation procedure (ensemble of five independently trained models, with epistemic uncertainty taken as the variance of their predictions; no Monte-Carlo dropout is used); (ii) the complete model architecture (terrain heightmap encoder followed by a task-conditioned MLP that receives both terrain features and commanded velocity); (iii) the training data distribution (10 000 procedurally generated terrain patches with controlled roughness, slope, and obstacle parameters); and (iv) the precise definition of feasibility error (fraction of predicted footholds that violate the robot’s dynamic constraints when executed in simulation, computed over a fixed horizon). These additions are accompanied by pseudocode and hyper-parameter tables. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical method with independent experimental validation

full rationale

The paper presents a data-driven approach to learning epistemic uncertainty in foothold predictions conditioned on terrain and motion inputs, then incorporating that uncertainty into costmaps for planning. No equations, derivations, or first-principles results are described that reduce to inputs by construction. The OOD detection and feasibility improvements are shown via separate simulation and real-world experiments rather than tautological fits or self-referential definitions. No load-bearing self-citations, uniqueness theorems, or ansatzes are invoked in the provided text. The derivation chain is self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard machine-learning assumptions about epistemic uncertainty estimation from neural networks trained on limited data; no free parameters, invented entities, or ad-hoc axioms are explicitly introduced in the abstract.

axioms (1)
  • domain assumption A neural network trained on limited terrain-motion data can produce epistemic uncertainty estimates that correlate with out-of-distribution regions
    Invoked to justify using uncertainty for OOD detection and costmap generation

pith-pipeline@v0.9.0 · 5479 in / 1187 out tokens · 35243 ms · 2026-05-09T19:41:58.833940+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

29 extracted references · 29 canonical work pages · 1 internal anchor

  1. [1]

    QuadPiPS: A Perception-informed Footstep Planner for Quadrupeds With Semantic Affordance Prediction

    Max Asselmeier, Ye Zhao, and Patricio A Vela. Steppability-informed quadrupedal contact planning through deep visual search heuristics.arXiv preprint arXiv:2501.00112, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    Advances in real-world applications for legged robots.Journal of Field Robotics, 35(8):1311–1326, 2018

    C Dario Bellicoso, Marko Bjelonic, Lorenz Wellhausen, Kai Holtmann, Fabian G¨ unther, Marco Tranzatto, Peter Fankhauser, and Marco Hutter. Advances in real-world applications for legged robots.Journal of Field Robotics, 35(8):1311–1326, 2018

    work page 2018

  3. [3]

    Single-shot foothold selection and constraint evaluation for quadruped locomotion

    Dominik Belter, Jakub Bednarek, Hsiu-Chin Lin, Guiyang Xin, and Michael Mistry. Single-shot foothold selection and constraint evaluation for quadruped locomotion. In2019 International Conference on Robotics and Automation (ICRA), pages 7441–7447. IEEE, 2019

    work page 2019

  4. [4]

    Evora: Deep evidential traversability learning for risk-aware off-road autonomy.IEEE Transactions on Robotics, 40:3756–3777, 2024

    Xiaoyi Cai, Siddharth Ancha, Lakshay Sharma, Philip R Osteen, Bernadette Bucher, Stephen Phillips, Jiuguang Wang, Michael Everett, Nicholas Roy, and Jonathan P How. Evora: Deep evidential traversability learning for risk-aware off-road autonomy.IEEE Transactions on Robotics, 40:3756–3777, 2024

    work page 2024

  5. [5]

    Pietra: Physics-informed evidential learning for traversing out-of- distribution terrain.IEEE Robotics and Automation Letters, 10(3):2359– 2366, 2025

    Xiaoyi Cai, James Queeney, Tong Xu, Aniket Datar, Chenhui Pan, Max Miller, Ashton Flather, Philip R Osteen, Nicholas Roy, Xuesu Xiao, et al. Pietra: Physics-informed evidential learning for traversing out-of- distribution terrain.IEEE Robotics and Automation Letters, 10(3):2359– 2366, 2025

    work page 2025

  6. [6]

    Estimating epistemic and aleatoric uncertainty with a single model.Advances in Neural Information Processing Systems, 37:109845–109870, 2024

    Matthew Chan, Maria Molina, and Chris Metzler. Estimating epistemic and aleatoric uncertainty with a single model.Advances in Neural Information Processing Systems, 37:109845–109870, 2024

    work page 2024

  7. [7]

    Dynamic path planning of a mobile robot adopting a costmap layer approach in ros2

    Pangcheng David Cen Cheng, Marina Indri, Fiorella Sibona, Matteo De Rose, and Gianluca Prato. Dynamic path planning of a mobile robot adopting a costmap layer approach in ros2. In2022 IEEE 27th International Conference on Emerging Technologies and Factory Automation (ETF A), pages 1–8. IEEE, 2022

    work page 2022

  8. [8]

    Foothold eval- uation criterion for dynamic transition feasibility for quadruped robots

    Luca Clemente, Octavio Villarreal, Angelo Bratta, Michele Focchi, Victor Barasuol, Giovanni Gerardo Muscolo, and Claudio Semini. Foothold eval- uation criterion for dynamic transition feasibility for quadruped robots. In 2022 International Conference on Robotics and Automation (ICRA), pages 4679–4685. IEEE, 2022. 12 Singh et al

    work page 2022

  9. [9]

    Fast algorithms to test robust static equilibrium for legged robots

    Andrea Del Prete, Steve Tonneau, and Nicolas Mansard. Fast algorithms to test robust static equilibrium for legged robots. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2016

    work page 2016

  10. [10]

    Pronav: Proprioceptive traversability estimation for legged robot navigation in outdoor environments.IEEE Robotics and Au- tomation Letters, 9(8):7190–7197, 2024

    Mohamed Elnoor, Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, and Dinesh Manocha. Pronav: Proprioceptive traversability estimation for legged robot navigation in outdoor environments.IEEE Robotics and Au- tomation Letters, 9(8):7190–7197, 2024

    work page 2024

  11. [11]

    Vital: Vision-based terrain-aware locomotion for legged robots.IEEE Transactions on Robotics, 39(2):885–904, 2022

    Shamel Fahmi, Victor Barasuol, Domingo Esteban, Octavio Villarreal, and Claudio Semini. Vital: Vision-based terrain-aware locomotion for legged robots.IEEE Transactions on Robotics, 39(2):885–904, 2022

    work page 2022

  12. [12]

    Task-driven out-of- distribution detection with statistical guarantees for robot learning

    Alec Farid, Sushant Veer, and Anirudha Majumdar. Task-driven out-of- distribution detection with statistical guarantees for robot learning. In Conference on Robot Learning, pages 970–980. PMLR, 2022

    work page 2022

  13. [13]

    Autonomous delivery robots and their potential impacts on urban freight energy consumption and emissions

    Miguel Figliozzi and Dylan Jennings. Autonomous delivery robots and their potential impacts on urban freight energy consumption and emissions. Transportation research procedia, 46:21–28, 2020

    work page 2020

  14. [14]

    Rloc: Terrain-aware legged locomotion using re- inforcement learning and optimal control.IEEE Transactions on Robotics, 38(5):2908–2927, 2022

    Siddhant Gangapurwala, Mathieu Geisert, Romeo Orsolino, Maurice Fal- lon, and Ioannis Havoutis. Rloc: Terrain-aware legged locomotion using re- inforcement learning and optimal control.IEEE Transactions on Robotics, 38(5):2908–2927, 2022

    work page 2022

  15. [15]

    Haptic inspection of planetary soils with legged robots.IEEE Robotics and Automation Letters, 4(2):1626–1632, 2019

    Hendrik Kolvenbach, Christian B¨ artschi, Lorenz Wellhausen, Ruben Grandia, and Marco Hutter. Haptic inspection of planetary soils with legged robots.IEEE Robotics and Automation Letters, 4(2):1626–1632, 2019

    work page 2019

  16. [16]

    Trajectory optimization for wheeled-legged quadrupedal robots driving in challenging terrain.IEEE Robotics and Au- tomation Letters, 5(3):4172–4179, 2020

    Vivian S Medeiros, Edo Jelavic, Marko Bjelonic, Roland Siegwart, Marco A Meggiolaro, and Marco Hutter. Trajectory optimization for wheeled-legged quadrupedal robots driving in challenging terrain.IEEE Robotics and Au- tomation Letters, 5(3):4172–4179, 2020

    work page 2020

  17. [17]

    Caldwell, and Claudio Semini

    Romeo Orsolino, Michele Focchi, Carlos Mastalli, Hongkai Dai, Darwin G. Caldwell, and Claudio Semini. Application of wrench-based feasibility anal- ysis to the online trajectory optimization of legged robots.IEEE Robotics and Automation Letters, 3(4):3363–3370, 2018

    work page 2018

  18. [18]

    Feasible region: An actuation-aware extension of the support region.IEEE Transactions on Robotics, 36(4):1239–1255, 2020

    Romeo Orsolino, Michele Focchi, St´ ephane Caron, Gennaro Raiola, Victor Barasuol, Darwin G Caldwell, and Claudio Semini. Feasible region: An actuation-aware extension of the support region.IEEE Transactions on Robotics, 36(4):1239–1255, 2020

    work page 2020

  19. [19]

    Rapid stability margin estimation for contact-rich locomotion

    Romeo Orsolino, Siddhant Gangapurwala, Oliwier Melon, Mathieu Geisert, Ioannis Havoutis, and Maurice Fallon. Rapid stability margin estimation for contact-rich locomotion. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8485–8492. IEEE, 2021

    work page 2021

  20. [20]

    Gram: Generalization in deep rl with a robust adaptation module.IEEE Robotics and Automation Letters, 2025

    James Queeney, Xiaoyi Cai, Alexander Schperberg, Radu Corcodel, Mouhacine Benosman, and Jonathan P How. Gram: Generalization in deep rl with a robust adaptation module.IEEE Robotics and Automation Letters, 2025

    work page 2025

  21. [21]

    Hierar- chical vision navigation system for quadruped robots with foothold adap- tation learning.Sensors, 23(11):5194, 2023

    Junli Ren, Yingru Dai, Bowen Liu, Pengwei Xie, and Guijin Wang. Hierar- chical vision navigation system for quadruped robots with foothold adap- tation learning.Sensors, 23(11):5194, 2023. Task-Conditioned Uncertainty Costmaps for Legged Locomotion 13

    work page 2023

  22. [22]

    Learning to walk in minutes using massively parallel deep reinforcement learning

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

    work page 2022

  23. [23]

    Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning.IEEE Robotics and Automation Letters, 5(2):3699– 3706, 2020

    Vassilios Tsounis, Mitja Alge, Joonho Lee, Farbod Farshidian, and Marco Hutter. Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning.IEEE Robotics and Automation Letters, 5(2):3699– 3706, 2020

    work page 2020

  24. [24]

    Rough terrain navigation for legged robots using reachability planning and template learning

    Lorenz Wellhausen and Marco Hutter. Rough terrain navigation for legged robots using reachability planning and template learning. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6914–6921. IEEE, 2021

    work page 2021

  25. [25]

    Where should i walk? predicting terrain properties from images via self-supervised learning.IEEE Robotics and Automation Letters, 4(2):1509–1516, 2019

    Lorenz Wellhausen, Alexey Dosovitskiy, Ren´ e Ranftl, Krzysztof Walas, Ce- sar Cadena, and Marco Hutter. Where should i walk? predicting terrain properties from images via self-supervised learning.IEEE Robotics and Automation Letters, 4(2):1509–1516, 2019

    work page 2019

  26. [26]

    Navigation planning for legged robots in challenging terrain

    Martin Wermelinger, P´ eter Fankhauser, Remo Diethelm, Philipp Kr¨ usi, Roland Siegwart, and Marco Hutter. Navigation planning for legged robots in challenging terrain. In2016 IEEE/RSJ International Conference on In- telligent Robots and Systems (IROS), pages 1184–1189. IEEE, 2016

    work page 2016

  27. [27]

    Foothold-based planning for legged robot autonomous navigation over uneven terrain.IEEE/ASME Transactions on Mechatronics, 2025

    Jiyu Yu, Dongqi Wang, Zhenghan Chen, Ci Chen, Rong Xiong, and Yue Wang. Foothold-based planning for legged robot autonomous navigation over uneven terrain.IEEE/ASME Transactions on Mechatronics, 2025

    work page 2025

  28. [28]

    Single-image footstep prediction for versa- tile legged locomotion

    Wuming Zhang and Kris Hauser. Single-image footstep prediction for versa- tile legged locomotion. In2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4407–4413. IEEE, 2018

    work page 2018

  29. [29]

    How should a robot assess risk? Towards an axiomatic theory of risk in robotics

    Anirudha Majumdar and Marco Pavone. How should a robot assess risk? Towards an axiomatic theory of risk in robotics. InRobotics Research: The 18th International Symposium (ISRR), pages 75–84. Springer, 2019

    work page 2019