Source-linked AI summary
Mine Tunnel Exploration using Multiple Quadrupedal Robots
Ian D. Miller, Fernando Cladera, Anthony Cowley, Shreyas S. Shivakumar, Elijah S. Lee, Laura Jarin-Lipschitz, Akhilesh Bhat, Neil Rodrigues, Alex Zhou, Avraham Cohen, Adarsh Kulkarni, James Laney, Camillo Jose Taylor, Vijay Kumar
TL;DR
Underground exploration must contend with difficult terrain and GNSS denial, while communication and endurance constraints limit conventional approaches. The paper presents autonomous multi-robot mine exploration, mapping, detection, and communication using legged quadrupeds; field missions included traversals up to 240m, though artifact localization within 5m was not achieved in one mine.
Problem
Underground mine exploration must handle challenging terrain, GNSS and communication denial, and sensory degradation while minimizing human intervention.
Source: §I · p. 1Method
The authors combine legged-robot exploration and planning with mapping, artifact detection, and communication systems for autonomous multi-robot mine exploration.
Source: §I · p. 2Results
A robot traversed 240m before slipping and falling on gravel, while another 165m traversal detected four artifacts and the detector had a 100% true positive rate across runs.
Source: B. Mine Experiments · p. 8Takeaways & Limitations
The authors demonstrate system feasibility in lab and field tests and conclude that the mapping, detection, and communication algorithms also apply to more general multi-robot exploration problems.
Source: §X · p. 8In the Experimental Mine, artifacts could not be localized within 5 meters because the corridor was spatially smooth and localization relied on LiDAR.
Source: B. Mine Experiments · p. 8
Abstract
from arXiv · showhide
Robotic exploration of underground environments is a particularly challenging problem due to communication, endurance, and traversability constraints which necessitate high degrees of autonomy and agility. These challenges are further exacerbated by the need to minimize human intervention for practical applications. While legged robots have the ability to traverse extremely challenging terrain, they also engender new challenges for planning, estimation, and control. In this work, we describe a fully autonomous system for multi-robot mine exploration and mapping using legged quadrupeds, as well as a distributed database mesh networking system for reporting data. In addition, we show results from the DARPA Subterranean Challenge (SubT) Tunnel Circuit demonstrating localization of artifacts after traversals of hundreds of meters. These experiments describe fully autonomous exploration of an unknown Global Navigation Satellite System (GNSS)-denied environment undertaken by legged robots.
I. INTRODUCTION
Underground mine exploration demands autonomy amid difficult terrain, GNSS and communication denial, and sensory degradation. The paper presents legged-robot exploration and planning alongside communication, detection, and mapping systems, and reports autonomous mine traversals.
I. INTRODUCTION
- Mines combine complex terrain, GNSS and communication denial, and sensory degradation, while practical deployment calls for robust, reliable automation with no human interaction after mission start.Limited bandwidth and connectivity also require onboard operation and careful transmission and storage of useful information.Source: §I · p. 1
- The system combines MAVs and legged quadrupeds because underground terrain varies, MAVs have flight-time, mission-length, and payload limits, and quadrupeds can carry larger sensor payloads for long-duration missions.Prior quadruped navigation work assumed GNSS and a pre-existing global map or manually defined path.Source: §I · pp. 1–2
- The authors present exploration and planning algorithms for autonomous exploration of underground tunnel environments.Detection and communication are described as platform-agnostic and used across a heterogeneous robot suite including quadrotors.Source: §I · p. 2
- The communication, detection, and mapping systems are designed to provide rapid situational awareness to a single operator in communication-challenged environments..Source: §I · p. 2
- Tests include lab and NIOSH mine experiments, with multiple autonomous legged-robot traversals of hundreds of meters in different mines.The experiments are described as long-duration and entirely autonomous.Source: §I · p. 2
II. SYSTEM OVERVIEW
The legged platform combines LiDAR, short-range depth sensors, visual-inertial tracking, and detection sensors with onboard compute. A mesh network connects the system, while the Xavier sends navigation commands to the gait controller.
A. Hardware Architecture
- The Nvidia Jetson AGX Xavier provides 8 ARM CPU cores, 512 GPU CUDA cores, and 16GB shared RAM for distributing software workloads.CPU/GPU distribution is particularly useful for mapping and image processing.Source: A. Hardware Architecture · p. 2
- The Ouster LiDAR supports navigation, planning, and mapping; two time-of-flight sensors provide short-range depth for local planning, and a RealSense camera estimates high-speed visual-inertial pose.A Zed Mini stereo camera, FLIR Boson IR camera, and Bluetooth radio support object detection; thermal sensing targets survivors and Bluetooth detects cell phones.Source: A. Hardware Architecture · p. 2
- A Rajant dual-band mesh system provides communication, and the Xavier sends high-level body-frame twist commands to the Vision 60 gait controller over UDP/Ethernet..Source: A. Hardware Architecture · p. 2
B. Software Architecture
The software runs largely onboard, separates local autonomy from global mapping, and uses a keyframe-based mapping pipeline. Local maps support tracking, while a pose graph and compressed keyframes support global reconstruction.
B. Software Architecture
- Separate ROS masters accommodate intermittent, unreliable underground communications, while the distributed database carries inter-robot data and separate channels handle base-station commands and telemetry.A launch manager starts, stops, and monitors subsystems, allowing operator checks without SSH or a direct ROS connection.Source: B. Software Architecture · p. 2
- Independent software subsystems separate local planning and control from the global mapper, so a mapper failure or divergence still leaves the robot able to choose feasible local directions.The architecture is shown in Fig. 3.Source: B. Software Architecture · p. 2, Figure 3 · p. 2
III. MAPPING
- The mapper links local maps or keyframes with a pose graph; panoramic maps record per-pixel depth, surface normal, and confidence, with free space implicit along measured rays.Panoramas retain representational power for maps gathered over limited translational excursions [12], [13].Source: §III · p. 2, Figure 4 · p. 3
- At 10Hz, LiDAR sweeps are de-rotated using a time-synchronized IMU, initialized from RealSense pose change, and registered to the active keyframe with projective ICP.The mapper typically runs at 50Hz on the Xavier; completed keyframes are pose-graph optimized, downsampled, and losslessly compressed as 16-bit PNGs for base-station point-cloud assembly.Source: §III · p. 3
IV. HIGH LEVEL NAVIGATION
The exploration system works locally, identifying candidate tunnels from LiDAR depth panoramas and selecting paths through them. It follows the tunnel closest to the current heading, using operator turn sequences to diversify robot routes.
IV. HIGH LEVEL NAVIGATION
- Tunnel exploration operates locally and must identify potential tunnels as well as select which tunnel to explore..Source: §IV · p. 3
A. Tunnel Detection
- A single LiDAR scan’s depth panorama yields tunnel candidates by thresholding distant connected components, tracking their centroids over time with an EKF, and suppressing candidates near projected recent poses.The projected pose history helps avoid inadvertent backtracking; Fig. 5 shows tracked tunnels and pose history.Source: A. Tunnel Detection · p. 3, Figure 5 · p. 3
B. Exploration Behavior
- The robot follows the tunnel closest to its heading, marks a tunnel untraversable when terrain cost is too high, and returns to base if no tunnels are visible..Source: B. Exploration Behavior · p. 3
- Operator-entered turn sequences direct different robots toward different routes; each detected tunnel on the specified side is selected and removes that turn from the queue.After the sequence is exhausted, the robot continues in its current tunnel until it cannot continue or exploration times out.Source: B. Exploration Behavior · p. 3
C. Return to Base Behavior
The robot returns to the entrance by following a path through the mapper’s topological graph, while local planning uses a gradient-based costmap to evaluate and generate trajectories. Terrain costs conservatively penalize unknown space and increase near tunnel walls.
C. Return to Base Behavior
- Breadth-first search finds a path from the current keyframe to the start, and tunnel-following planning guides the robot along it or directly toward the next keyframe.When the desired direction differs sufficiently from a detected tunnel, the robot ignores detected tunnels; the behavior could also navigate to other keyframes, though it was used only to return to the start.Source: C. Return to Base Behavior · pp. 3–4
V. LOCAL PLANNING AND CONTROL
- Local planning centers a 2D costmap on the robot and models traversability using terrain-gradient magnitude; Fig. 6 shows the pointcloud-to-pruned-trajectory pipeline.Source: §V · p. 4, Figure 6 · p. 4
A. Terrain Characterization
- LiDAR scans become ground-plane heightmaps whose gradients represent terrain steepness and roughness; the robot assumes motion can be integrated in SE(2) without significant error.Unknown gradient regions receive infinite cost, and gradient-based costs filter out absolute elevation.Source: A. Terrain Characterization · p. 4
- Configuration-space generation models the robot as a 2D rectangle R(x, y, θ); GPU conversion takes ~10 ms on Xavier, and a 1D Gaussian blur along y raises costs near tunnel walls, pushing the robot toward the tunnel center.Unknown cells are assigned infinite cost, treating unknown space as untraversable.Source: A. Terrain Characterization · p. 4
B. Terrain Planning
The planner searches configuration space for a trajectory toward a desired direction rather than selecting individual footsteps or targeting a fixed goal state. It combines trajectory penalties with hard traversability constraints, then prunes the path to smooth it.
B. Terrain Planning
- The planner evaluates robot-footprint traversability in configuration space instead of optimizing individual footsteps, and uses a desired direction rather than an explicit goal state.Its cost combines traversability and trajectory-length penalties, reflecting legged robots’ continuous rather than binary traversability.Source: B. Terrain Planning · p. 4
- The hybrid objective combines trajectory cost with direction cost while restricting admissible paths to those that avoid cells above a configuration-space cost threshold.Source: B. Terrain Planning · p. 5
- The cost penalizes sidestepping and backward motion, because sidestepping is less stable and the robot’s sensors cannot instantly see behind it.The desired-direction cost assumes the high-level explorer supplies generally open local directions; if path cost is too high, the tunnel tracker selects a new direction.Source: B. Terrain Planning · p. 5
- Dijkstra’s search stops after reaching a point sufficiently far in the desired direction, then recursive pruning removes cost-neutral waypoints and linearly interpolates to smooth the trajectory.Source: B. Terrain Planning · p. 5
C. Control
The generated path is time-parameterized at fixed linear and angular velocities, then tracked with a proportional controller.
C. Control
- At each time step, a proportional controller commands a twist based on the difference between the planned trajectory state and the planner’s actual state.The trajectory is time-parameterized assuming fixed linear and angular velocities.Source: C. Control · p. 5
VI. OBJECT DETECTION
Artifact detection combines semantic segmentation and depth to estimate artifact positions, while Bluetooth scans provide a separate cellphone-detection channel. The system clusters repeated visual detections for reporting and logs Bluetooth detections with robot pose.
A. Vision Based Detection
- ERFNet semantic segmentation detects the SubT artifact classes, including fire extinguishers, backpacks, hand-drills, cellphones, and survivors.The network was selected for compact, efficient inference on the onboard compute hardware.Source: A. Vision Based Detection · p. 5
- The 3,416-image, 1280 × 720 training dataset combines human-labeled images from varied environments and sensors with cave and mine video frames, some augmented with pasted artifacts.Fig. 7 shows artifact examples with per-pixel training annotations.Source: A. Vision Based Detection · p. 5, Figure 7 · p. 5
- Training used half-resolution images, split train and test data across different environments, and selected the model with the lowest test mean intersection over union for relevant classes.Source: A. Vision Based Detection · p. 5
- Segmentation and stereo depth yield artifact 3D positions, which are transformed to the nearest mapping keyframe and clustered across detections by class and spatial proximity.Clusters receive confidence from the fraction of detections agreeing within a time window; outliers and possible false positives are still sent to the operator for review.Source: A. Vision Based Detection · p. 5
B. Cellphone Detection
- A Bluetooth module scans for enabled devices and records each device’s MAC ID and signal strength, pushing a detection when its signal strengthens and tagging it with robot pose.Source: B. Cellphone Detection · pp. 5–6
VII. DISTRIBUTED DATABASE
The distributed database shares robot data across an intermittently connected network, while the strategy treats communication as optional and lets each robot act on its own information. A state machine governs exploration and return, with operator controls for pausing or stopping robots.
VII. DISTRIBUTED DATABASE
- Each network node shares information through a distributed database, synchronizing new or changed messages when link quality is sufficient and limiting each robot to writing its own data.Robots can read data published by other nodes, allowing them to use peers’ information without complex consistency procedures.Source: §VII · p. 6
VIII. STRATEGY, HUMAN INTERFACE
- The system treats communication as a luxury: each robot bases its behavior on its own observations and information, with high-level state machines triggering actions.The strategy is designed for low-bandwidth, intermittent subterranean links.Source: §VIII · p. 6
A. State Machine
- START sets the exploration time limit and turn specification; EXPLORE searches for tunnels, and exceeding the time limit sends the robot to MOVEBASE.STOP and ESTOP can be manually triggered to pause or sit down the robot.Source: A. State Machine · p. 6, Figure 8 · p. 6
- For Tunnel Circuit, robots were deployed one at a time with decreasing time limits to mitigate inter-robot interference and maintain mesh-network communication.Inter-robot collaboration was proposed as future work that could improve performance.Source: A. State Machine · p. 6
VIII. STRATEGY, HUMAN INTERFACE
- The interface figure shows a robot-status monitor alongside each robot’s artifact listing and an artifact viewer.The figure caption identifies these as elements of one user-interface instance.Source: Figure 9 · p. 6
B. User Interface
The operator interface combines robot status, artifact detections, and map views. In a controlled motion-capture mission, the LiDAR-based mapper outperformed RealSense tracking, with reported errors around or below a meter after operator filtering.
B. User Interface
- The interface presents robot status, artifact detections with images and bounding boxes, a global point-cloud reconstruction, and top-down paths and artifact locations.The operator can review artifact images before sending them to DARPA.Source: B. User Interface · p. 6
A. Lab Experiment
- After a roughly 30 m traversal, total trajectory error was about a meter, or roughly 3%; detected objects were localized within a meter.Some backpack detections were misclassified as fire extinguishers, but a human operator could filter them.Source: A. Lab Experiment · p. 7
- Within a meter, filtered artifact errors and absolute trajectory error remained; LiDAR mapping outperformed RealSense tracking on ATE and RPE.RPE used a fixed two-second window.Source: A. Lab Experiment · p. 7, Table I · p. 7
B. Mine Experiments
Mine trials demonstrated artifact detection and reporting over long traversals, but artifact localization varied by mine and long-duration missions ended in falls. The authors present the system as applicable to broader multi-robot exploration and identify mapping and control as areas for further work.
B. Mine Experiments
- About half of 10–15 missions had sensor failures, and all long-duration missions ended with falls on gravel, slippery surfaces, or surface dips.The local planner was overconfident, and its gradient-based traversability metric missed hazards such as gravel and wet metal.Source: B. Mine Experiments · p. 7
- In the Safety Research Mine, the robot traveled 240 m before slipping and falling on gravel, after more than 2,000 steps over terrain including train tracks.A 165 m traverse in Fig. 10 represented over 1,000 successful steps.Source: B. Mine Experiments · p. 8
- The detector achieved a 100% true positive rate; a cell phone was reported within 5 m after about 100 m in the Safety Research Mine, while Experimental Mine artifacts could not be localized within 5 m.The Experimental Mine traversal was 165 m, with four artifacts correctly detected and some false positives filtered by the operator.Source: B. Mine Experiments · p. 8
X. CONCLUSION
- The authors present exploration, planning, mapping, and detection technologies for autonomous legged-robot mine exploration, and state that mapping, detection, and communication also apply to broader multi-robot exploration.They identify further mapping and control work to improve robustness and accuracy.Source: §X · p. 8