Heterogeneous Ground and Air Platforms, Homogeneous Sensing: Team CSIRO Data61's Approach to the DARPA Subterranean Challenge

1 Introduction

• SubT requires a single operator to direct robots through dangerous, communications-limited underground terrain and find as many artefacts as possible within one hour.The challenge includes increasing difficulty from distance to urban obstacles and ambiguous natural-cave traversability.Source: §1 · p. 2
• The team combines platforms with different terrain and access strengths while sharing lidar-and-camera sensing, Wildcat multi-agent SLAM, and 3D frontier exploration.Large tracked robots traverse rough cave terrain; smaller UGVs pass narrower doorways and stairs, while UAVs provide rapid overviews and access to promising structures unreachable from the ground.Source: §1 · p. 2
• Agents independently build global maps from peer-shared SLAM frames, while operators can confirm detections, launch UAVs, and influence task priorities.The paper describes operator control ranging from single-agent teleoperation to consensus-based task allocation with no operator input.Source: §1 · pp. 2–3

1.1 Related Work

• Prior work spans lidar robot-team mapping, staged UAV/UGV awareness, simulated rescue, multi-operator UGV tasks, disaster-response goals, and phased 3D UGV/UAV exploration.The cited examples include base-station map merging, operator-base planning, and separate ground and aerial mapping phases.Source: §1.1 · pp. 3–4

1.2 Other SubT Teams

• Other SubT entries ranged from Pluto’s topological fixed-duration exploration and Explorer’s submap ledger to CoSTAR’s risk-aware planning, Cerberus’s graph planners, Marble’s reactive navigation, and NCTU’s blimp.The paper contrasts these approaches with its emphasis on a common understanding across agents for coordination beyond operator-base range.Source: §1.2 · pp. 4–5

1.3 Contributions

• CatPack and Hovermap provide equivalent sensing for ground and air platforms, while CatPack enables reuse of the perception and autonomy stack across ground robots.The sensing packs use equivalent sensors and common SLAM and 3D frontier logic.Source: §1.3 · p. 5, Figure 1 · p. 5
• Wildcat supplies real-time SLAM, and shared frames let agents exchange artefacts, frontiers, and traversability data; the Urban Event gate example won most accurate artefact detection.The reported detection involved one robot detecting the gate and reporting it to others through the shared SLAM solution.Source: §1.3 · pp. 5–6
• The system also contributes peer-to-peer topometric navigation, disruption-tolerant communications, and multi-agent task allocation that accounts for platform capabilities and operator priorities.The topometric map combines local navigation-cost data and segments it using superpixel methods, allowing an agent to navigate to frontiers or artefacts found by another.Source: §1.3 · p. 6

2 System Description

• The fleet composition changed by event, while modular sensing and shared power and emergency-stop systems reduced the complexity of maintaining multiple platforms.In the Urban and Cave circuits, large tracked robots carried UAVs for in-field launch, allowing them to pass through narrow openings and conserving UAV battery life.Source: §2 · p. 6

2.1 CatPack

• The custom CatPack provides a common UGV sensing payload, with aligned cameras and lidar supporting both mapping and near-field navigation.Its cameras form near-continuous 360° horizontal coverage; the thermal cameras had not been used because they were not considered necessary for artefact detection.Source: §2.1 · pp. 6–8

2.2 UGV Platforms

• UGVs share the CatPack and a common power, emergency-stop, communications, navigation, and control architecture, while platform choice balances capability, size, complexity, payload, robustness, and cost.The emergence of rugged tracked and quadruped robots prompted reconsideration of the custom hexapod effort.Source: §2.2 · pp. 8–9

2.2.1 SuperDroid LT2-F

• The SuperDroid LT2-F handled concrete floors and narrow stairwells, but low clearance and track derailments made rough, granular terrain difficult and maintenance-intensive.These limitations prompted investment in a larger tracked platform and development of a smaller tracked robot.Source: §2.2.1 · p. 9

2.2.2 BIA5 All Terrain Robot

• The 90 kg BIA5 ATR proved reliable and robust, traversed rough terrain, and could carry communications gear and a UAV, but its size prevented entry through narrow doorways.The larger ATR and smaller UGVs therefore provided complementary capabilities.Source: §2.2.2 · pp. 9–10

2.2.3 Ghost Robotics Vision60

• Electrical and power reliability concerns led the team to deprioritize the Vision60 for Urban and Cave, despite later hardware iterations showing promise on rough terrain, narrow gaps, and stairs.The team planned to reconsider it for the final event after further locomotion testing.Source: §2.2.3 · p. 10

2.2.4 CSIRO Hexapod

• The Bruce hexapod demonstrated rough-terrain traversal at 0.3 m/s and flat-ground speeds up to 0.5 m/s, but actuator thermal cut-off constrained sustained top-speed operation.The team later halted development, diverting resources to other priorities because of the effort needed to achieve desired reliability and capability.Source: §2.2.4 · p. 11

2.2.5 CSIRO Dynamic Tracked Robot

• The DTR was designed to bring ATR-like step-climbing to a smaller footprint, with track geometry enabling stair climbing without flippers.The figure shows a CAD drawing beside a 180 mm step for reference.Source: §2.2.5 · p. 11, Figure 7 · p. 11

2.3 UAV Hardware

• The DJI M210 carries an Emesent Hovermap payload, a custom gimballed camera system, and a Rajant radio.The commercial Hovermap payload runs Emesent’s autonomy stack with added integration and perception functionality.Source: §2.3 · p. 12
• A three-axis gimbal helps limit motion blur from UAV movement and longer exposures in poorly lit subterranean environments.The UAV uses a single camera system on the gimbal for artefact detection and localisation.Source: §2.3 · p. 12

2.4 Communication Systems

• The early Ubiquiti mesh used self-erecting, self-powered drop nodes, but in-band communication halved bandwidth at each hop and meshing proved unreliable.The drop mechanism and self-erection worked as designed, despite communication limitations.Source: §2.4 · p. 12
• A three-module ESP-32 node alternated frequency bands between neighbours, but transmissions overloaded receiving modules on the same node.Source: §2.4 · p. 12
• The team shifted from developing a pervasive fixed network toward greater agent autonomy, allowing robots to operate beyond communication range and return to report data.Source: §2.4 · p. 12
• Rajant ES-1 nodes served ground robots and drop nodes, while UAVs used single-band DX-2 nodes; the team retained drop nodes to seek persistent operator connectivity.The ES-1 nodes were dual-band.Source: §2.4 · p. 13
• Agents store data in Mule and share it with nearby agents and the base when connected; disconnected agents return to communication range independently or through task allocation.Source: §2.4 · p. 13
• After few nodes were dropped at Urban, cave-event UGVs carried one drop node and a hardwired ES-1; a redesigned dropper reduced footprint and obstacle-collision likelihood.The cave dropper no longer hung off the robot’s back, improving manoeuvrability in tight spaces.Source: §2.4 · p. 13, Figure 10 · p. 14

• SubT scoring requires artefacts to be detected, correctly classified, and localised within five metres of their surveyed locations.Source: §3 · p. 14
• Agents must autonomously explore complex environments, traverse challenging terrain, and build accurate maps for navigation and artefact localisation.Source: §3 · p. 14
• Each fleet robot is equipped to explore, map, locate artefacts, and communicate results to other agents and the base station.Source: §3 · p. 14

3.1 Perception

• UGVs and UAVs primarily use RGB cameras to detect artefacts and projected lidar points to localise them; UGVs also use signal analysers and gas sensors.The additional UGV sensors detect cell phones and high CO2 concentrations.Source: §3.1 · p. 14
• The operator reviews, confirms, and reports detections through a GUI that displays a sortable list, global-map location, and associated image.Source: §3.1 · p. 14

3.1.1 Camera Based Detection

• The annotated artefact dataset contains more than 13,000 images from nine environments, with poorly lit scenes, varied poses, occlusions, multiple artefacts, and negative examples.It includes 3,412 images without artefacts.Source: §3.1.1 · p. 14
• Detector training was iteratively updated by annotating frames when practice-event performance was poor and retraining the model.The authors say this manual inspection approach indicates detector performance in a new environment.Source: §3.1.1 · p. 15
• Detections below artefact-specific posterior-probability thresholds are removed to keep the operator’s detection load manageable.Thresholds are calculated in advance using approximately 45–60-minute datasets from agents operating in anticipated similar environments.Source: §3.1.1 · p. 15
• UGV detection uses DeNet at 12 Hz across four cameras, while the UAV uses MobileNet V2 SSD 300 at 3 Hz on its Intel CPU.The UGV implementation uses TensorRT; the UAV implementation uses OpenVINO and training data containing false positives from aircraft parts.Source: §3.1.1 · p. 15

• An artefact location tracker was added to each UGV before Urban after novel environments produced unmanageable false detections, mostly repeat detections.Source: UGV Artefact Localisation and Tracking · p. 15
• Lidar points projected inside each detected bounding box are weighted toward nearer points to estimate artefact location; detections beyond 20 m are ignored.Source: UGV Artefact Localisation and Tracking · p. 15
• The tracker models known artefact positions with isotropic Gaussians and new artefacts uniformly, creating a new track beyond 1 m from a known artefact.Tracks are dropped after more than 30 seconds without redetection, and detections are linked to agent trajectories for later multi-agent SLAM correction.Source: UGV Artefact Localisation and Tracking · p. 15
• 0.46 m RMS error was achieved across nine artefacts reported from autonomous UGV camera detections at Urban; one report had 0.22 m error, the lowest of any team.Source: UGV Artefact Localisation and Tracking · p. 16

UAV Artefact localisation and Gimbal Pointing Autonomy

• A stabilised gimbal camera decouples perception from flight autonomy, allowing views into tight, un-flyable spaces and side passages away from flight goals.Source: UAV Artefact localisation and Gimbal… · p. 16
• The gimbal tracks the camera viewpoint with greatest expected utility, updated on each exposure, to inspect side passages and newly unoccluded spaces.Expected utility depends on the camera sensor model and prior utility of visible space.Source: UAV Artefact localisation and Gimbal… · p. 16
• At the Urban event, UAV artefact reports had 2.34 m RMS error and were less accurate than the UGV system; detections were not tracked onboard.The UAV projected bounding-box centroids into the odometry frame using ranges from a spherical map of occupied voxels, averaging points in the nearest 1 m range bin.Source: UAV Artefact localisation and Gimbal… · p. 16

3.1.2 Additional UGV Sensing

• UGVs detected CO2 through within-room concentration spikes and cell phones through Bluetooth and WiFi signal strength, reporting reception locations for operator interpretation.CO2 spikes avoided intermittent false negatives from simple thresholding under changing atmospheric pressure, altitude, and temperature.Source: §3.1.2 · p. 16
• Five of 15 Urban event artefact reports came from additional sensing, including three cell phone detections missed by RGB cameras.Source: §3.1.2 · p. 16

3.2 Odometry

• Wildcat odometry optimises trajectory poses in a five-second sliding window, sampled every 0.1 seconds, using inertial disagreement and lidar matching errors.The approach is based on Bosse et al. (2012).Source: §3.2 · pp. 16–17
• Lidar points are voxelised into 40 cm surfels every 0.5 seconds, then nested into multi-resolution grids up to 3.2 m with an offset finest grid.Each surfel is represented by the mean and covariance of its points, which depend on the corresponding trajectory.Source: §3.2 · p. 17
• Surfel correspondences use a plane-to-plane error from the difference in surfel means projected onto the combined covariance's smallest-eigenvalue eigenvector.A tree-based k-nearest-neighbour lookup considers position, normal, and time; trajectory steps use iterative re-weighted least squares with Cauchy weights for outlier robustness.Source: §3.2 · p. 17
• Points leaving the five-second window enter consolidated multi-resolution surfels in six-second frames generated every five seconds from data 5–11 seconds old.These frames provide stable local coordinate systems for locating artefacts and referencing exploration and navigation data over time and across agents.Source: §3.2 · p. 17

3.3 Exploration

• Frontiers mark boundaries between known and unknown space, where loop closures undermine a single global map and 2D occupancy maps do not capture subterranean 3D structure.Source: §3.3 · p. 17
• Visible-space point clouds and maximum-range sphere points identify frontiers; long edges in the triangulated visible mesh can also indicate openings.The watertight mesh of visible points is termed a viewpoint, while segmented frontiers represent free-space boundaries against unknown space.Source: §3.3 · p. 17
• Frontier processing repeats after each metre of agent motion, comparing old and new viewpoints to remove frontier vertices now visible from another viewpoint.Frontiers and viewpoints are stored in the latest Wildcat frame coordinate system; Figure 11 illustrates the processing stages.Source: §3.3 · p. 17, Figure 11 · p. 18
• Exploration progressed from segmented-frontier goals in the Tunnel Circuit to autonomous multi-agent exploration in the Urban Circuit, with observer locations selected using estimated visible volume behind frontiers.For cave testing, candidate observer locations used traversability analysis from local and global navigation.Source: §3.3 · p. 18
• Agents share frontiers but not viewpoints, discounting rewards near previously visited observer locations; this lowers communications needs but reduces performance versus sharing complete viewpoint data.The same frontier solution runs on UGVs and UAVs, enabled by homogeneous sensing and mapping, though frontier selection differs.Source: §3.3 · p. 18

3.4 UGV Local Navigation

• Local navigation combines mapping, deliberative planning, and reactive behaviours to help UGVs traverse mud, rubble, puddles, and concrete while handling negative obstacles, slopes, steps, overhangs, and narrow passages.The modules use terrain slope, visibility, and vehicle orientation to recognise and react around unobserved areas while minimising collisions.Source: §3.4 · pp. 18–19
• OHM integrates SLAM-corrected point clouds into a 3D occupancy map and per-cell ground-height map, including best-case virtual surfaces for obscured areas.Figure 13 distinguishes observed possibly traversable cells, observed non-traversable cells, and unobserved virtual surfaces; Figure 12 shows OHM among adjacent navigation components.Source: §3.4 · p. 19, Figure 12 · p. 19, Figure 13 · p. 20
• Fatal cells are identified from height changes between nearby cells, with resilience to rough terrain and small holes, and are excluded from the robot footprint's planned path.Source: §3.4 · p. 19
• Hybrid A* evaluates terrain and costmap cells under the robot footprint, fatally costing intersections with fatal cells and terrain too steep for safe traversal.Heightmap cells also estimate robot pitch and roll to increase the cost of difficult terrain.Source: §3.4 · p. 19
• The highest-priority admissible behaviour supplies the velocity command, with separate behaviours for path following, obstacle avoidance when planning fails, and preventing tip-over.Independent behaviours can be modified or added with minimal impact on other behaviours.Source: §3.4 · pp. 19–20

3.5 UAV navigation

• An Autonomy Level 2 manager coordinates local and global planning, receives operator tasks, and interfaces with lower-level autonomy.The modular architecture simplifies continued improvement of higher-level autonomy.Source: §3.5 · p. 20
• The UAV supports exploration and 3D, 2D, and planar waypoints, ranging from exact spatial goals to general directions of travel.A 3D waypoint fixes position; a 2D waypoint fixes X and Y but leaves height free; a planar waypoint specifies a plane without a precise goal point.Source: §3.5 · p. 20
• UAV exploration reuses the fleet’s frontier method, sending sufficiently large frontiers to local navigation at approximately 0.5 Hz.Frontiers are biased toward large, nearby options in the UAV’s current direction; using direct metric distance rather than global-map reasoning about obstacles has not proved a significant drawback during limited-duration operations.Source: §3.5 · p. 20
• At 2 m/s, the UAV has travelled more than 600 m underground and returned autonomously in one flight when wires are no smaller than 4 mm.It reliably operates in a 3 m tunnel at 2 m/s; it operates at 1 m/s to optimise artefact detection and to avoid a vertically dangling 1 mm wire.Source: §3.5 · pp. 20–21

4 Multi-Agent Mapping and Data Sharing

• Each agent computes a global map in a decentralised manner from its currently available subset of the data collected by all agents.Agents share mapping, intent, artefact, and status data; the base station is also treated as an agent, and agents periodically exchange manifests and request missing information.Source: §4.1 · p. 21, §4 · p. 21

4.1.1 SLAM Frames

• Wildcat integrates lidar and IMU data into shared SLAM frames, then builds an Atlas Graph to solve the global alignment of those frames.Frames contain surfels and an IMU integration estimate rather than raw sensor data, reducing bandwidth; Figure 14 illustrates successive odometry portions and their associated features.Source: §4.1 · p. 21, §4.1.1 · p. 22, Figure 14 · p. 22

4.1 Multi-Agent SLAM

• Wildcat’s Tunnel Circuit SLAM error was < 0.05% of distance travelled; the distributed solution supported coordination while disconnected and contributed to the Urban event’s most accurate artefact detection.The authors report Wildcat was, to their knowledge, the only multi-agent SLAM solution run in a distributed manner at the most recent Urban Circuit.Source: §4.1 · p. 21

4.1.2 Atlas Pose Graph Optimisation

• The hierarchical pose graph uses frame poses and odometry, geometric-alignment, and loop-closure constraints; marginalising redundant frames makes graph complexity grow with distance travelled rather than time.Figure 15 depicts the pose-graph representations and constraints. Agent maps are not guaranteed to be metrically or topologically consistent when frame-arrival order differs, though common starting areas yield sufficiently similar solutions for practical purposes.Source: §4.1 · p. 21, §4.1.2 · p. 22, Figure 15 · p. 24
• Current broadcasting sends every frame to every agent, ensuring access to the complete data set but scaling bandwidth linearly with each added agent.Stationary agents also publish redundant information; delta-style communications or partial map transfers are future work.Source: §4.1.2 · p. 23

4.2 Multi-Agent UGV Global Navigation

• UGV global navigation uses a sparse topometric graph built from all agents’ cost maps and Wildcat poses, linking traversable subgraphs across frames.Cost maps are bundled by Wildcat frame; the graph uses 2D submaps and superpixel-derived subgraphs, with traversable inter-subgraph edges determined by submap intersections. Figure 16 illustrates this construction.Source: §4.2 · pp. 23, 25, Figure 16 · p. 25

4.3 Data Sharing

• Mule replaces base-station-brokered sharing with peer-to-peer, disruption-tolerant ROS bridging, synchronising persistent data through manifests and pull requests.Volatile topics use best-effort delivery, while persistent topics such as Wildcat frames and artefact detections are stored on disk for later transfer; agents can use synchronisation status to decide whether to return to base.Source: §4.3 · pp. 25–26

5 Human Robot Team Decision Making

• Cave testing used a general task-allocation framework for autonomous robot coordination while allowing operators to guide team behaviour.The stated end goal is for operators to focus on team-level mission decisions rather than individual robot behaviour.Source: §5 · p. 26

5.1 Finite State Automaton Design

• Nested finite-state automata compose primitive behaviours into higher-level actions, such as sequencing navigation, node deployment, and moving away from the dropped node.The executive constructs and executes an FSA from operator command lists, ticking it at 1 Hz; primary behaviours include goto, explore, sync, drop-comms-node, and launch-UAV.Source: §5.1 · pp. 26–27

5.2 Multiple-Robot Task Allocation

• Market-based task allocation has agents bid on tasks using expected reward, while ordered task bundles account for the reward change from adding a task and avoid full combinatorial optimisation.Core tasks are explore, drop-comms-node, and sync-data; allocation must adapt to new or completed tasks, communication loss, and changing team composition.Source: §5.2 · p. 27
• Frontier rewards were combined with priority levels, whose thresholds made the team explore the main trunk first and then fill in details.Reward tuning is challenging; priority provides a way to order tasks without estimating reward adjustments.Source: §5.2 · p. 27
• Auction consensus is decentralised: agents broadcast task bids and use shared rules to choose winners, but unreliable communication can leave agents with conflicting task states.When communication returns, agents share task sets to reach consensus; competing agents are generally close enough to communicate, which mitigates many difficulties. Figure 17 illustrates the consensus process.Source: §5.2 · pp. 27–28, Figure 17 · p. 28

5.3 Incorporating Operator Guidance

• Operators can place manual tasks or use 3D priority regions to adjust task rewards and priorities, including for selected task types or agents.Regions can prioritise useful areas or block areas considered unproductive or dangerous; Figure 19 shows the priority-region interface and artefact-review panel.Source: §5.3 · pp. 28–29, Figure 19 · p. 29

6 Discussion

• Across the evaluations, testing led the team to prioritize robotic autonomy and coordinated fleets over further communications development and reliance on one operator for higher-level guidance.The deployed communications, UGV fleet, multi-agent autonomy, and operating concepts changed substantially between events.Source: §6 · p. 30

6.1 STIX

• At STIX, waypoint-only control, operator workload, and a Miami Tunnel constriction 30 m inside the entrance prompted work on higher autonomy and marsupial UAV launches.The constriction prevented further UAV progress, while rough terrain and beaching helped motivate acquiring a larger UGV.Source: §6 · p. 30, §6.1 · p. 31

6.2 Tunnel Circuit

• During the tunnel event, 11 of 80 artefacts were correctly detected and located; at 215 m from staging, RMS errors over distance travelled were 0.63% for odometry, 0.22% for real-time atlas, and 0.025% for offline global estimation.Source: §6.2 · pp. 32–33

6.3 Urban Circuit

• Ground robots traversed 7,902 m across four runs; narrow doorways were often missed because of traversability-analysis resolution, while data synchronization clustered robots near the entry, later addressed by MRTA’s frontier-cluster task allocation.In one instance, an operator identified a vertical shaft, and a UAV descended, autonomously traversed the lower level, and returned; its trajectory is shown in Figures 24 and 25.Source: §6.3 · p. 34
• 15 of 80 Urban Circuit artefacts were correctly detected and localised; 12 were detected by UGVs, two by both UGVs and UAVs, and one by the operator inspecting the map.Correctly reported artefacts had a median error of 0.58 m, and the team received the most accurate detection award for an error of 0.22 m.Source: §6.3 · pp. 34, 36

6.4 Cave Circuit

• Cave navigation identified traversable ground well, but under-modelled terrain costs favored short, difficult routes; insufficient waypoint risk modelling sometimes left agents stuck or tasks temporarily blocked.Local navigation was judged more reliable than operator teleoperation in the challenging terrain.Source: §6.4 · pp. 37–38
• In the cave event, the UAV operated as a scout and covered more terrain than ground agents; the course was smaller than Tunnel and Urban because it was limited to safely accessible areas.On the beta course, the green ground agent traversed 231 m and the UAV traversed 400 m; four of ten artefacts were detected.Source: §6.4 · p. 38
• 16 of 44 cave-event artefacts were correctly detected and located; 14 were detected by UGVs and two by the operator inspecting the map.Three additional artefacts were visible but too distant to detect, and one was missed because detector thresholds were too high.Source: §6.4 · p. 38

7 Conclusions

• Platform heterogeneity increased course coverage: large UGVs carried UAVs, UAVs reached shafts or otherwise inaccessible terrain, and smaller UGVs accessed narrow areas.Smaller UGVs also served as additional communication relays.Source: §7 · p. 38
• At each event, the agents commanded with the highest available autonomy were consistently the most effective in points scored.Source: §7 · p. 38
• Homogeneous sensing and Wildcat SLAM supported shared global maps, which underpinned information sharing, navigation coordination, artefact sharing, and autonomous task allocation.Agents gained coordination through operator situational awareness and multi-robot task allocation.Source: §7 · pp. 38, 40
• Across events, the team shifted effort away from pervasive communications and complex legged robots toward simpler, robust platforms and greater autonomy.Source: §7 · p. 38
• In cave testing, alpha’s steep, narrow, rocky segment remained unexplored; beta was nearly fully explored, but its large cavern was incomplete.Source: Figure 28 · p. 40
• Follow-on work targets terrain-aware frontier selection and navigation, better exploration prioritization, and autonomous behaviors for narrow openings, recovery, and close-proximity cooperation.Source: §7 · pp. 40–41
• Other planned work includes improving operator point-cloud awareness and UAV perception, and developing smart nodes to store and later relay passing agents’ data.Source: §7 · p. 41
• The team planned to extend its focus on autonomy, coordination, and platform heterogeneity to larger environments at the September 2021 SubT final.Platform work also included improving robustness and stability, such as avoiding drop-node mounts that increase footprint and reduce maximum slope.Source: §7 · p. 41