Subterranean Agent Robotics

Updated 23 June 2026

Subterranean agents are robotic systems designed to explore and map underground environments where GPS, lighting, and communications are severely limited.
They integrate multi-modal sensors with advanced SLAM, planning, and behavior synthesis to navigate complex, hazardous terrains.
Field deployments and competitions like the DARPA SubT Challenge validate their efficiency in real-world, industrial and rescue applications.

A subterranean agent is an autonomous or semi-autonomous robotic system designed for exploration, mapping, inspection, and intervention within underground environments such as mines, tunnels, caves, and urban subsurfaces. These agents operate under severe constraints, including the absence of GPS, degraded or denied communications, low to zero ambient light, and geometrically complex or hazardous terrain. Subterranean agents can be ground-based, aerial, or composed of heterogeneous multi-agent fleets, integrating advanced perception, localization, planning, decision architectures, and resilient networking. The state-of-the-art in subterranean agents is exemplified by contributions from the DARPA Subterranean Challenge and extensive real-world deployments in industrial mining operations (Dahlquist et al., 17 Jan 2025).

1. System Architectures and Hardware Modalities

Subterranean agents span ground robots (UGVs), aerial robots (UAVs), and, in advanced systems, hybrid and legged forms or marsupial architectures. Exemplary platforms include:

Aerial Quadrotor Agents: Quadrotors with 3D LiDAR + IMU for high-rate odometry (e.g., FAST-LIO), RGB-D or stereo cameras for inspection and object detection, onboard computers (Intel NUC), and PX4 or Pixhawk flight controllers. These agents execute full local autonomy stacks with receding-horizon planning (e.g., D $^*_+$ , NMPC, or APF), BT-driven mission execution, and reactive obstacle avoidance. Fielded systems often support auction-based multi-agent collaboration and mesh networking (Dahlquist et al., 17 Jan 2025, Tranzatto et al., 2022, Petris et al., 2022).
Ground Robotics: Tracked, wheeled, or legged UGVs (e.g., ANYmal C quadrupeds, BIA5 ATRs) integrating spinning/solid-state LiDAR, multiple RGB cameras, IMUs, and high-end compute resources (Jetson Xavier, AMD Ryzen CPUs). These support high-end SLAM, traversability analysis, advanced planning (RRT*, PRM*, ScanPlan), and deep learning for artifact identification (Tranzatto et al., 2022, Hudson et al., 2021, Tranzatto et al., 2022).
Specialized Modalities: Blimps for resource-constrained and collision-tolerant long-duration scouting (Duckiefloat), collision-tolerant multirotors (RMF-Owl, Gagarin) for narrow, debris-prone shafts, and relay-robot chains (U-Chain) for maintaining ad hoc comm links (Huang et al., 2019, Petris et al., 2022, Laclau et al., 2020).
Network and Operator Layers: Mesh Wi-Fi (802.11ax/802.11s, Ubiquiti, Rajant, Mobilicom), droppable “breadcrumb” nodes, and operator GUIs for task management and 3D visualization. Some systems employ distributed, peer-to-peer SLAM coordination without central servers (Dahlquist et al., 17 Jan 2025, Hudson et al., 2021).

2. Perception, Localization, and Mapping

Perception systems leverage multi-modal, multi-sensor fusion for robust state estimation and artifact detection:

SLAM Algorithms: Tightly coupled LiDAR-inertial SLAM (FAST-LIO, LOAM, LIO-SAM, CompSLAM, Wildcat), often operating at 10–100 Hz. Loop closures are performed via scan-matching, pose-graph optimization (GTSAM/iSAM2), and robust kernels (e.g., Huber, Cauchy, PCM/ICM outlier rejection). Volumetric mapping uses occupancy voxels (OctoMap, Voxblox) with resolutions commonly around 0.15–0.2 m (Dahlquist et al., 17 Jan 2025, Tranzatto et al., 2022, Ebadi et al., 2020, Agha et al., 2021).
Map Fusion and Decentralization: Multi-agent systems merge local OctoMaps using diff-maps (~100× bandwidth reduction), Bayesian updates, and coordinate alignment (Horn’s absolute orientation via Leica prisms for sub-degree accuracy) (Ohradzansky et al., 2021). Peer-to-peer map fusion employs covariance intersection for redundant overlapping cells (Agha et al., 2021).
Visual and Semantic Perception: Deep-learning-based object detection (YOLOv3, MobileNet-SSD) on RGB(-D)/thermal imagery, fused with depth estimates via back-projection and Bayesian hypothesis filtering. Event cameras are fused with LiDAR for robust detection under harsh lighting (Saucedo et al., 2023, Patel et al., 2022).
Traversability and Risk Mapping: 2.5D/3D elevation and semantic maps, with traversability scores computed from slope, roughness, and occupancy. Agents may fuse traversability with CVaR-based uncertainty models for risk-aware planning (Agha et al., 2021, Tranzatto et al., 2022).

3. Planning, Task Allocation, and Behavior Synthesis

Auction-Based Task Allocation: Tasks (e.g., inspection of specified locations) are entered by human operators into a 3D GUI and pooled. Each agent computes cost bids—typically the length or risk cost of their D $^*_+$ -planned path to each task—and submits them to a ROS-based auctioneer. The auctioneer solves an integer linear program to maximize task-assignment profit while ensuring exclusivity constraints:

$x^* = \arg\max_{x\in\{0,1\}^{n_a \times n_t}} \sum_{k=1}^{n_a} \sum_{j=1}^{n_t} \rho_{k,j}\, x_{k,j} \ \text{s.t. } \sum_k x_{k,j}\leq 1,\ \sum_j x_{k,j}\leq 1$

Assignments are broadcast; new operator tasks are inserted in real-time, triggering auction recomputation with in-progress tasks held out (Dahlquist et al., 17 Jan 2025).

Behavior Tree Synthesis: Back-chaining from operator goals using a library of action primitives ( $\mathcal L$ ) with specified preconditions and postconditions. BTs are synthesized recursively via sequences/fallbacks to ensure modularity, verifiability, and reactivity. New tasks or capabilities are integrated by extending the primitive library (Dahlquist et al., 17 Jan 2025).
Exploration and Navigation: Hybrid local/global planning—sampling-based (RRT*, PRM*) and graph-based (GBPlanner2, ScanPlan, Information RoadMap) approaches. Frontiers are prioritized via information gain over unknown voxels, discounted by traversal cost or risk. Topological exploration (SER segmentation) leverages dense maps and keyframe contribution graphs to enable efficient global planning and minimize redundant backtracking (Kim et al., 2023, Tranzatto et al., 2022, Agha et al., 2021).

4. Communication, Networking, and Coordination

Mesh Comms Infrastructure: Battery-powered Wi-Fi 6 mesh (Ubiquiti U6-Pro), droppable relays (Rajant DX2, Mobilicom), and peer-to-peer ad hoc networks (Meshmerize, Nimbro, udp_mesh). Data exchange protocols prioritize essential packets (artifact reports, telemetry) and utilize diff-maps for efficient map updates. Robustness against signal loss is achieved through buffer-on-loss, with autonomous operation for >5 minutes in comm-denied zones (Dahlquist et al., 17 Jan 2025, Rouček et al., 2021, Ohradzansky et al., 2021).
Relay Chains and U-Chain Protocol: Signal-based self-organization maintains a UAV relay chain with minimal relays, ensuring that each link’s RSSI is balanced and above thresholds. Distributed Kalman-filtered RSSI and local movement enable convergence to nearly optimal relay positioning under severe constraints (no SLAM/GPS, only RSSI+optic flow) (Laclau et al., 2020).
Operator Interface: GUIs present live agent positions, task states, and mapping data; allow real-time task insertion and teleoperation when needed. Autonomy stacks are designed to minimize operator workload, with demonstrated low intervention rates in field trials (Dahlquist et al., 17 Jan 2025, Biggie et al., 2023).

5. Robustness, Collision Tolerance, and Failure Recovery

Collision Tolerance: Use of non-rigid or foam-laminate structural elements (e.g., RMF-Owl carbon-foam chassis, Duckiefloat blimp envelope), elastic mountings, and passive mechanical compliance to absorb energy in physical impacts. Autonomy stacks treat collision as a recoverable event—rapid deviation from planned vs. actual pose triggers replanning and recovery actions rather than catastrophic mode switches (Huang et al., 2019, Petris et al., 2022).
Resource and Power Management: Strict payload optimization (e.g., trading LiDAR for lighter stereo vision when mass is constrained), dynamic power management (duty-cycled active illumination, low-thrust “buoyancy park” modes), and hot-swap battery interfaces extend mission lifetimes (Huang et al., 2019).
Failure Detection and Recovery: Embedded fault-detection and safety mechanisms—visual odometry “feature drop” thresholds, collision-acceleration triggers, and automated fallback to low-bandwidth communications or human recovery. Decision cost functions aggregate multiple failure indicators to trigger mode switches (“request human assistance”/return to base) (Huang et al., 2019).
Comms Degradations: Incrementally deployed, manually or autonomously placed relay nodes, with buffer-based artifact and map reporting in communications-degraded regimes. Some architectures support peer-to-peer synchronization on rendezvous, eliminating single points of failure (Hudson et al., 2021, Rouček et al., 2021).

6. Field Deployments, Metrics, and Lessons Learned

Industrial Validation: Large-scale underground mining deployments (LKAB mine, 500 m depth, 200 m mapped), with three aerial agents executing 7 inspection tasks in 311 s. All tasks completed with <0.2 m localization drift and average path-tracking error of 0.3 m (Dahlquist et al., 17 Jan 2025).
DARPA SubT Challenge: Multiple teams (e.g., CERBERUS, CSIRO, MARBLE, CoSTAR, CTU-CRAS-NORLAB) demonstrated kilometer-scale traversal, mapping accuracy (RPE/APEs <0.5 m and <2°), low artifact localization errors (<0.3 m), efficient multi-robot exploration (18–23 artifacts scored per mission), and robust performance under sensor degradation and communication loss (Tranzatto et al., 2022, Hudson et al., 2021, Biggie et al., 2023, Agha et al., 2021, Rouček et al., 2021).
Quantitative Metrics:
- SLAM/localization drift: <0.2–0.5 m over 200+ m traversed.
- End-to-end proposal–detection latency: 1.5 s (LiDAR-guided PTZ).
- Mission duration: 5–60 min (depending on platform/power).
- Map coverage: 49–79% of ground-truth volumes.
- Exploration rate improvement: +62% via keyframe-based topological planning (SER) over baseline (Kim et al., 2023).
Key Lessons:
- Modular autonomy stacks and homogeneous sensing accelerate development, particularly for heterogeneous fleets (Hudson et al., 2021).
- Mesh stability and careful placement are decisive in large undergrounds; resilience to single-node failures is essential.
- Automated behavior tree synthesis and rapid map-diff-based multi-agent coordination are empirically superior for scaling system robustness and operator throughput (Dahlquist et al., 17 Jan 2025).
- Extensive underground field testing is indispensable for closing the sim-to-real gap.

7. Research Directions and Limitations

Scaling Task Allocation: Fully centralized auction cycles can bottleneck at agent count ≥5. Partial decentralization or hierarchical auction layers are potential routes for future research (Dahlquist et al., 17 Jan 2025).
Integration of Advanced Risk Models: Existing path planners utilize D $^*_+$ risk-aware metrics; integrating environmental hazard data (gas, seismic) is an open extension.
Specialization vs. Generalization: Systems optimized for narrow tubes/tunnels (U-Chain, blimps) may not generalize to open caverns; modular stacks and topological planners show stronger adaptability (Laclau et al., 2020, Kim et al., 2023).
Communication Bandwidth and Robustness: Map compression, event-driven updates, and opportunistic data mule architectures remain active research areas for bandwidth-limited domains (Agha et al., 2021, Hudson et al., 2021).
Learning-Enabled Autonomy: Online adaptation of traversability models (neural nets/GPs), risk bounds (CVaR), and uncertainty management are under exploration for increased resilience in novel domains (Agha et al., 2021).

Subterranean agents represent the leading edge of robust, autonomous robotics under extreme operational constraints, fusing multi-modal perception, resilient planning, adaptive networking, and modularized control for safety-critical industrial and search-and-rescue applications in large-scale underground environments (Dahlquist et al., 17 Jan 2025, Tranzatto et al., 2022, Kim et al., 2023).