MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment".
Dev: Here's the gist of the paper. Underground mines present extreme conditions for autonomous robot navigation: GPS is denied, lighting is degraded, and tunnel topology is loop-rich and non-convex MineXplore presents an open-source MuJoCo-based navigation benchmark derived from the Leung et al.
Rosa: 2017 Chilean underground copper mine dataset This environment reconstructs a 104,423 m2 tunnel network through an six-stage contour-to-MJCF pipeline incorporating octagonal wall cross-sections, LiDAR-sourced jagged wall geometry, three terrain friction zones, a global 5◦ incline, and periodic spot lighting A single-agent PPO baseline trained via RLlib across five independent random seeds achieves a best rolling coverage of 88.89% ± 1.74% (3 of 5 seeds reaching the 90% coverage target), confirming that MineXplore supports stable and reproducible policy learning under realistic underground sensing and topology Environment Construction MineXplore's geometry is derived from a real operational mine The compilation pipeline involves six stages, starting with source inputs like the published 2D survey floor-plan and LiDAR point cloud logs Stage 1 involves Source Data and Scale Calibration, where pixel-to-metric scale is calibrated against a 100 m scale bar yielding 1.36 px/m Stage 2 focuses on 2D Geometry Extraction using OpenCV contour detection to define the tunnel footprint and rock islands Stage 3 handles 3D Compilation, which includes polygon triangulation, extrusion to a nominal height of 3.5 m, and V-HACD convex decomposition into 1,186 collision geoms Stage 4 models Surface Wall Geometry by using a two-layer system comprising an 8-sided polygonal base layer and a LiDAR-sourced jagged overlay Realism Features Implementation The environment incorporates several features to replicate the physical conditions of the Chilean copper mine These realism features include: 1.
Dev: Terrain incline, where the entire MJCF worldbody is rotated 5◦ around the Y-axis, producing an effective elevation gain of 5.2 m across the entrance region 2.
Rosa: Terrain friction zones, which are implemented through three spatially fixed floor box geoms replicating dry (1.00), muddy (0.55), and waterlogged (0.25) zones 3.
Paper summary: Rosa: So we’re wrapping up our look at "MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment". This paper shows how they took a real Chilean copper mine dataset and compiled it into a MuJoCo environment that's ready for reinforcement learning one <ref:2606.04569#pg1>.
Dev: They spent their time detailing the six stages of this compilation pipeline, from initial 2D survey inputs to the final validated MJCF model, which is key because it shows how to ground simulation in real-world data three <ref:2606.04569#pg2>.
Taro: What’s significant about the conclusion is that they validate this geometric fidelity at an intersection over union of zero point nine five three eight against the source survey map three <ref:2606.04569#pg2>.
Rosa: They also report surface texture similarity scores of seventy nine point four percent across six structural dimensions, which gives us a sense of how accurate their reconstruction is compared to the original survey one <ref:2606.04569#pg1>.
Dev: The overall finding is that this environment supports stable and reproducible policy learning under realistic underground sensing and topology, even though the agent achieved a mean final coverage of eighty-four point five three percent across five independent seeds one <ref:2606.04569#pg1>.
Taro: The implication for autonomy is that having a simulation grounded in real production-mine geometry closes a gap for testing agents in these extreme underground environments, which is what this benchmark aims to do one <ref:2606.04569#pg1,grounded in real production-mine geometry>.
Rosa: So, the title "MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment" points to this being the first environment compiled from a real operational mine inside MuJoCo/MJX one <ref:2606.04569#pg1,MineXplore: An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground>.
Dev: It’s positioning it as a resource for open-source reinforcement learning exploration in environments where GPS is denied, lighting is degraded, and the tunnel topology is loop-rich and non-convex one <ref:2606.04569#pg1,tunnel topology is loop-rich and non-convex>.
Taro: The limitation they mention is that this work focuses on local observations and how the environment supports stable exploration learning under those conditions one <ref:2606.04569#pg1>.
Conclusion: Rosa: So, we're wrapping up our look at MineXplore, which is basically an open-source benchmark for testing how robots navigate underground mines without any GPS signal or good lighting one.
Dev: Yeah, it’s about taking a real Chilean copper mine and turning it into a MuJoCo environment that lets us study reinforcement learning in these tough conditions one.
Rosa: The authors are using this whole six-stage pipeline to build the model, starting from simple 2D floor plans and LiDAR data three <ref:2606.04569#pg2>.
Dev: They focus on making sure the geometry is solid, which they say means a very high intersection over union score of zero point nine five three eight against the original survey map three.
Rosa: That’s a lot of detail going into making sure the simulation actually looks like a real mine tunnel, including those jagged wall surfaces and different floor friction zones one.
Dev: The performance results show that their single-agent policy baseline got an aggregate best rolling coverage of eighty-eight point eight nine percent across five different random seeds one.
Rosa: So, what does this mean for the world? It means we finally have a simulation environment that’s built from actual operational mine data, closing the gap between lab testing and real-world underground autonomy one.
Dev: It supports stable policy learning under these realistic sensing and topology conditions, which is a big step for how we train robots for those kinds of jobs one.
Rosa: The authors point out that this work shows how local observations can lead to stable exploration learning in these challenging spaces one.
Dev: Their next steps involve using the full three dimensional LiDAR logs to make the geometry even better, and adding procedural changes to make the environment more diverse one.
Rosa: It’s an important foundation for future work on autonomous exploration in environments where you don't have perfect information, which brings us to how agents handle things when they misbehave next.
Abhishek S, Badrikanath Praharaj, Sreeram M.V.
BuildMachineLabs
cs.RO
Submitted: 2026-06-03
Updated: 2026-11-09
Comments: 7 pages,11 figures, This paper was accepted at Xplore: Cross-Disciplinary Aspects of Exploration in Robotics, Reinforcement Learning, and Search, a workshop held in conjunction with the 2026 IEEE International Conference on Robotics and Automation (ICRA 2026)
Code: https://github.com/osrf/subt
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 82/100
The gist: Underground mines present extreme conditions for autonomous robot navigation: GPS is denied, lighting is degraded…
Key concepts
- MineXplore
- A simulation environment built in MuJoCo that mimics a real underground mine. It reconstructs a complex 104,423 m2 tunnel network using LiDAR data and includes features like varying floor friction, periodic spot lighting, and a global incline to simulate extreme conditions.
- Contour-to-MJCF Pipeline
- A six-stage process used to convert raw mine survey data (like 2D plans and LiDAR point clouds) into the 3D geometry format required by MuJoCo. This pipeline handles everything from scaling measurements to creating complex, jagged wall surfaces for accurate simulation.
- Reward Structure
- The system uses a specific reward function designed to encourage active exploration rather than just reaching a goal. The agent is rewarded based on newly discovered cells ($\Delta c \cdot g$) and penalized for collisions, ensuring the learned policy focuses on mapping unknown areas of the tunnel.
- PPO Baseline
- Proximal Policy Optimization (PPO) is a specific type of reinforcement learning algorithm used to train the robot agent. It was used here to test if an RL policy could successfully navigate and explore the complex MineXplore environment without catastrophic failures.
Terminology
Summary
Underground mines present extreme conditions for autonomous robot navigation: GPS is denied, lighting is degraded, and tunnel topology is loop-rich and non-convex MineXplore presents an open-source MuJoCo-based navigation benchmark derived from the Leung et al. p. 0 com
Abstract—Underground mines present extreme conditions for
autonomous robot navigation: GPS is denied, lighting is degraded, and tunnel topology is loop-rich and non-convex. Simulation benchmarks grounded in real production-mine geometry and
compatible with GPU-accelerated learning pipelines do not yet
exist in the open-source ecosystem. 2017 Chilean underground copper mine dataset
[1]. The environment reconstructs a 104,423 m2
tunnel network
through an six-stage contour-to-MJCF pipeline incorporating octagonal wall cross-sections, LiDAR-sourced jagged wall geometry,
three terrain friction zones, a global 5◦
incline, and periodic
spot lighting.Where the paper says it
Where the paper says it
Environment Construction
MineXplore's geometry is derived from a real operational mine The compilation pipeline involves six stages, starting with source inputs like the published 2D survey floor-plan and LiDAR point cloud logs Stage 1 involves Source Data and Scale Calibration, where pixel-to-metric scale is calibrated against a 100 m scale bar yielding 1.36 px/m Stage 2 focuses on 2D Geometry Extraction using OpenCV contour detection to define the tunnel footprint and rock islands Stage 3 handles 3D Compilation, which includes polygon triangulation, extrusion to a nominal height of 3. p. 1 We binarised the survey floor-plan and extracted the tunnel footprint using OpenCV contour detection. The largest
external contour defines the outer tunnel boundary; interior
contours define rock islands and pillars.Where the paper says it
Realism Features Implementation
The environment incorporates several features to replicate the physical conditions of the Chilean copper mine These realism features include:
-
Terrain incline, where the entire MJCF worldbody is rotated 5◦ around the Y-axis, producing an effective elevation gain of 5. p. 3
2 m across the entrance regionWhere the paper says it
The entire MJCF worldbody is rotated 5◦ (0. 0872665 rad) around the Y -axis, tilting walls, floor, rocks, and lights uniformly as a rigid body.
-
Terrain friction zones, which are implemented through three spatially fixed floor box geoms replicating dry (1. p. 3
00), muddy (0. p. 3Where the paper says it
Terrain friction zones. Three spatially fixed floor box geoms partition the tunnel along its east–west axis to replicate mine water seepage: a dry western zone (friction 1.
55), and waterlogged (0. p. 3Where the paper says it
Three spatially fixed floor box geoms partition the tunnel along its east–west axis to replicate mine water seepage: a dry western zone (friction 1. 00), a muddy central zone (friction 0.
25) zonesWhere the paper says it
00), a muddy central zone (friction 0. 55), and a waterlogged eastern zone (friction 0.
-
Periodic lighting, featuring thirty-three MuJoCo spot lights placed along the tunnel centreline at 20 m intervals
Navigation and Reward Structure
The agent interacts with MineXplore using a Gymnasium-style reset/step interface The observation space is 19-dimensional, including LiDAR range readings from a simulated 2D scanner covering the full 360◦ field of view The action space is a continuous 2D box A = [-1, 1]2 mapped to differential-drive commands where speed and angular velocity are scaled by terrain type The reward structure is designed for exploration, defined as r = ∆c · g − 1[∆c = 0] · 0.002 + M − 1[collision] · 15. p. 4 Reward Structure
The per-step reward r is:
r = ∆c · g − 1[∆c = 0] · 0. 002 + M − 1[collision] · 15.Where the paper says it
Validation and Performance
The geometric fidelity of the compiled model is validated through two axes: navigability and geometric consistency For navigability, every pixel marked navigable in the source floor-plan maps to an unobstructed column in the MJCF model Geometric consistency is confirmed by a strict Intersection over Union (IoU) of 0.9538 against the source survey map The surface texture similarity scores 79. p. 0 Geometric fidelity is validated at an Intersection
over Union (IoU) of 0. 9538 against the source survey map, and
surface texture similarity scores 79. Six structural dimensions
were scored on a per-axis similarity metric (Fig. 8): wall
jaggedness (97. Per-axis surface-texture similarity comparison between the compiled
MineXplore model and the real Chilean mine LiDAR point cloud. All
six dimensions score above 50%; dominant wavelength (99. 7), and confirmed navigability using a
single-agent PPO baseline that achieves a best rolling coverage of 88. 74% with a mean final coverage of
84. 95%, giving an
aggregate best rolling coverage of 88. Coverage fraction over training (mean ± 1 SD across five seeds;
dotted line: 90% target).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Future Work
MineXplore is positioned as the first environment compiled from a real operational mine inside MuJoCo/MJX, closing the gap for a real-mine-grounded, Gymnasium-compatible navigation environment The work suggests that the environment supports stable exploration learning under local observations Future work includes full elevation-aware geometry from the Leung et al. 3D LiDAR logs and procedural cross-section perturbation for environment diversity The authors thank BuildMachineLabs for providing the opportunity, research direction, and resources that supported this work The authors also thank Leung et al. p. 5 3D LiDAR logs,
LiDAR-replay validation against the original dataset sweeps,
procedural cross-section perturbation for environment diversity, and broader multi-agent and cross-algorithm baselines. ACKNOWLEDGMENT
The authors thank BuildMachineLabs for providing the
opportunity, research direction, and resources that supported
this work.Where the paper says it
REFERENCES
[1] K. Y. K. Leung, D. Luhr, H. Houshiar, F. p. 5 REFERENCES
[1] K. Houshiar, F. Houshiar, F. Inostroza, D. Inostroza, D. Borrmann,
M. Borrmann,
M. Nuchter, and J. Ruiz-del-Solar, “Chilean underground ¨
mine dataset,” The International Journal of Robotics Research, vol. 16–23, 2017.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[2] M. Tranzatto, T. p. 5 1177/0278364916679497. Tranzatto, T. Tranzatto, T. Dharmadhikari, L. Dharmadhikari, L. Bernreiter, M. Bernreiter, M. Kulkarni,
F. Kulkarni,
F. Mascarich, O. Mascarich, O. Andersson, S. Andersson, S. Khattak, M.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[3] K Ebadi L Bernreiter H Biggie G Catt Y Chang et al Present and future of SLAM in extreme environments The IEEE Transactions on Robotics vol 40 pp 936–959 2024
[4] Open Robotics DARPA SubT Virtual Testbed 2021 Online Available: https://github.com/osrf/subt
[5] D Perille A Truong X Xiao and P Stone Benchmarking metric ground navigation in Proc IEEE Int Symp Safety Security and Rescue Robotics SSRR 2020
[6] V Makoviychuk L Wawrzyniak Y Guo M Lu K Storey M Macklin D Hoeller N Rudin A Allshire A Handa and G State Isaac Gym High performance GPU based physics simulation for robot learning in NeurIPS Datasets and Benchmarks Track 2021
[7] K Zakka B Tabanpour Q Liao M Haiderbhai S Holt J Y Luo A Allshire E Frey K Sreenath L Kahrs C Sferrazza Y Tassa and P Abbeel MuJoCo Playground arXiv preprint arXiv:2502.08844 2025
[8] Y Gao and K Awuah-Offei Navigation in underground mine environments A simulation framework for quadruped robots in Proc IEEE Int Conf Automation Science and Engineering CASE 2025 pp 1464–1469
[9] J Schulman F Wolski P Dhariwal A Radford and O Klimov Proximal Policy Optimization Algorithms arXiv preprint arXiv:1707.06347 2017
[10] M Towers A Kwiatkowski J Terry J U Balis G De Cola T Deleu M Goulao A Kallinteris M Krimmel A KG R Perez-Vicente A Pierre S Schulhoff J J Tai H Tan and O G Younis Gymnasium A Standard Interface for Reinforcement Learning Environments arXiv preprint arXiv:2407.17032 2024
[11] E Liang R Liaw R Nishihara P Moritz R Fox K Goldberg J Gonzalez M Jordan and I Stoica RLlib Abstractions for Distributed Reinforcement Learning in Proc 35th Int Conf Machine Learning ICML 2018 pp 3053–3062
[12] E Todorov T Erez and Y Tassa MuJoCo A physics engine for model-based control in Proc IEEE/RSJ Int Conf Intelligent Robots and Systems IROS 2012 pp 5026–5033
[13] B Yamauchi A frontier-based approach for autonomous exploration in Proc IEEE Int Symp Computational Intelligence in Robotics and Automation CIRA 1997 pp 146–151
[14] A Y Ng D Harada and S Russell Policy invariance under reward transformations Theory and application to reward shaping in Proc 16th Int Conf Machine Learning ICML 1999 pp 278–287
[15] M G Bellemare S Srinivasan G Ostrovski T Schaul D Saxton and R Munos Unifying count-based exploration and intrinsic motivation in Advances in Neural Information Processing Systems NeurIPS vol 29 2016
[16] MineXplore An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment arXiv:2606.04569 v1 Jun 3 2024 p. 0 04569v1 [cs. RO] 3 Jun 2026Where the paper says it
[17] K Y K Leung D Luhr H Houshiar F Inostroza D Borrmann M Adams A Nuchter and J Ruiz-del-Solar Chilean underground mine dataset The International Journal of Robotics Research vol 36 no 1 pp 16–23 2017 DOI: 10.1177/0278364916679497
[18] M Tranzatto T Miki M Dharmadhikari L Bernreiter M Kulkarni F Mascarich O Andersson S Khattak M Hutter R Siegwart and K Alexis CERBERUS in the DARPA Subterranean Challenge Science Robotics vol 7 no 66 p eabp9742 2022
[19] K Ebadi L Bernreiter H Biggie G Catt Y Chang et al Present and future of SLAM in extreme environments The IEEE Transactions on Robotics vol 40 pp 936–959 2024
[20] Open Robotics DARPA SubT Virtual Testbed 2021 Online Available: https://github.com/osrf/subt
[21] D Perille A Truong X Xiao and P Stone Benchmarking metric ground navigation in Proc IEEE Int Symp Safety Security and Rescue Robotics SSRR 2020
[2606.04569] MineXplore An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment arXiv:2606. p. 0 MineXplore: An Open-Source Reinforcement
Learning Exploration Benchmark for GNSS-Denied
Underground Environment
Abhishek S*
BuildMachineLabs
abhishekss6363@gmail. com
Badrikanath Praharaj*
BuildMachineLabs
badrikanath.Where the paper says it
[27] K Y K Leung D Luhr H Houshiar F Inostroza D Borrmann M Adams A Nuchter and J Ruiz-del-Solar Chilean underground mine dataset The International Journal of Robotics Research vol 36 no 1 pp 16–23 2017 DOI: 10.1177/0278364916679497
[28] M Tranzatto T Miki M Dharmadhikari L Bernreiter M Kulkarni F Mascarich O Andersson S Khattak M Hutter R Siegwart and K Alexis CERBERUS in the DARPA Subterranean Challenge Science Robotics vol 7 no 66 p eabp9742 2022
[16] MineXplore An Open-Source Reinforcement Learning Exploration Benchmark for GNSS-Denied Underground Environment arXiv:2606.04569 v1 Jun 3 2024
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Better exploration under local observations: The PPO baseline trained on MineXplore confirms that
MineXplore supports stable and reproducible policy learning under realistic underground sensing and topology,
specifically achieving abest rolling coverage of 88.89% ± 1.74%.
-
Grounded simulation for real-world deployment: The environment allows for training policies on a
real production-mine geometry
using theLeung et al. 2017 Chilean underground mine dataset,
which enables sim-to-real transfer to actual mining scenarios. -
Robust navigation under varying terrain conditions: The system can handle complex traction degradation by modeling
three terrain friction zones
(dry, muddy, waterlogged) and coupling these withactuation-level scaling
to model thefull traction degradation across ground conditions.
-
High-fidelity sensor simulation: The observation space provides a detailed input for the agent with a
19-dimensional observation vector,
including LiDAR range readings from 16 beams, which is normalized and augmented with Gaussian noise, mimicking realistic underground sensing. -
Curriculum-based exploration strategy: The reward structure implements milestone bonuses at
25%, 50%, and 75%
coverage thresholds to provide acurriculum-style waypoints that prevent the policy stalling at local coverage maxima during early training.
Abstract
Underground mines present extreme conditions for autonomous robot navigation: GPS is denied, lighting is degraded, and tunnel topology is loop-rich and non-convex. Simulation benchmarks grounded in real production-mine geometry and compatible with GPU-accelerated learning pipelines do not yet exist in the open-source ecosystem. We present MineXplore, an open-source MuJoCo-based navigation benchmark derived from the Leung et al. 2017 Chilean underground copper mine dataset. The environment reconstructs a 104,423 sq.m tunnel network through an six-stage contour-to-MJCF pipeline incorporating octagonal wall cross-sections, LiDAR-sourced jagged wall geometry, three terrain friction zones, a global 5 degree incline, and periodic spot lighting. Geometric fidelity is validated at an Intersection over Union (IoU) of 0.9538 against the source survey map, and surface texture similarity scores 79.4% across six structural dimensions. A single-agent PPO baseline trained via RLlib across five independent random seeds achieves a best rolling coverage of 88.89% (3 of 5 seeds reaching the 90% coverage target), confirming that MineXplore supports stable and reproducible policy learning under realistic underground sensing and topology.
Sources
- MuJoCo Playground
- Proximal Policy Optimization Algorithms
- Gymnasium: A Standard Interface for Reinforcement Learning Environments
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation