PaperlayerView exact revision on arXiv

Source-linked AI summary

CERBERUS: Autonomous Legged and Aerial Robotic Exploration in the Tunnel and Urban Circuits of the DARPA Subterranean Challenge

Marco Tranzatto, Frank Mascarich, Lukas Bernreiter, Carolina Godinho, Marco Camurri, Shehryar Khattak, Tung Dang, Victor Reijgwart, Johannes Loeje, David Wisth, Samuel Zimmermann, Huan Nguyen, Marius Fehr, Lukas Solanka, Russell Buchanan, Marko Bjelonic, Nikhil Khedekar, Mathieu Valceschini, Fabian Jenelten, Mihir Dharmadhikari, Timon Homberger, Paolo De Petris, Lorenz Wellhausen, Mihir Kulkarni, Takahiro Miki, Satchel Hirsch, Markus Montenegro, Christos Papachristos, Fabian Tresoldi, Jan Carius, Giorgio Valsecchi, Joonho Lee, Konrad Meyer, Xiangyu Wu, Juan Nieto, Andy Smith, Marco Hutter, Roland Siegwart, Mark Mueller, Maurice Fallon, Kostas Alexis

arXiv:2201.07067v1cs.RO

TL;DR

Subterranean robots must operate amid rough terrain, limited visibility, poor communications, and difficult access. CERBERUS combines legged and aerial robots with multimodal perception, mapping, planning, and deployable communications. Its report describes Tunnel and Urban Circuit deployments, the system’s contributions to resilience, and lessons and limitations from those experiences.

  • Problem

    Underground settings pose autonomy challenges including rough dynamic terrain, limited visibility, poor communication, and hard access.

    Source: §3 · p. 5

  • Method

    CERBERUS combines quadrupeds and aerial scouts with multimodal perception, multi-robot map optimization, and autonomous exploration capabilities.

    Source: §1 · p. 2

  • Results

    The report presents Tunnel and Urban Circuit performance and concludes that CERBERUS facilitated certain levels of resilience in visually degraded, large-scale, narrow, complex underground settings.

    Source: §13 · pp. 43–44

  • Takeaways & Limitations

    The report offers deployment lessons and open-source research intended to contribute to resilient robotic autonomy in extreme underground environments.

    Source: §13 · p. 44

    Tunnel deployments exposed elevation-map issues from foot slippage or corrupted RealSense depth data, prompting sensor and controller changes before the Urban Circuit.

    Source: §10.4.1 · p. 35

Abstract

from arXiv · show

Autonomous exploration of subterranean environments constitutes a major frontier for robotic systems as underground settings present key challenges that can render robot autonomy hard to achieve. This has motivated the DARPA Subterranean Challenge, where teams of robots search for objects of interest in various underground environments. In response, the CERBERUS system-of-systems is presented as a unified strategy towards subterranean exploration using legged and flying robots. As primary robots, ANYmal quadruped systems are deployed considering their endurance and potential to traverse challenging terrain. For aerial robots, both conventional and collision-tolerant multirotors are utilized to explore spaces too narrow or otherwise unreachable by ground systems. Anticipating degraded sensing conditions, a complementary multi-modal sensor fusion approach utilizing camera, LiDAR, and inertial data for resilient robot pose estimation is proposed. Individual robot pose estimates are refined by a centralized multi-robot map optimization approach to improve the reported location accuracy of detected objects of interest in the DARPA-defined coordinate frame. Furthermore, a unified exploration path planning policy is presented to facilitate the autonomous operation of both legged and aerial robots in complex underground networks. Finally, to enable communication between the robots and the base station, CERBERUS utilizes a ground rover with a high-gain antenna and an optical fiber connection to the base station, alongside breadcrumbing of wireless nodes by our legged robots. We report results from the CERBERUS system-of-systems deployment at the DARPA Subterranean Challenge Tunnel and Urban Circuits, along with the current limitations and the lessons learned for the benefit of the community.

1 Introduction

CERBERUS presents a heterogeneous legged-and-aerial system for autonomous subterranean exploration, combining multimodal perception, multi-robot mapping, path planning, and deployable communications. The paper reports its Tunnel and Urban Circuit deployments and the lessons learned.

1 Introduction

  • The paper details CERBERUS technology and reports Tunnel and Urban Circuit performance, mapping, autonomy, object detections, challenges, and lessons learned.It also identifies selected open-source software packages.Source: §1 · pp. 2–4
  • CERBERUS combines ANYmal quadrupeds, conventional and collision-tolerant aerial scouts, and the Armadillo rover for subterranean exploration and communications.ANYmal provides endurance and can deploy wireless nodes; the rover carries a high-gain antenna while maintaining an optical-fiber connection to the Base Station.Source: §1 · p. 2
  • LiDAR, camera, and inertial cues support resilient localization, while centralized Base Station map optimization refines detected-object positions.The introduction describes map optimization as combining individual robot estimates to improve reported positions.Source: §1 · p. 2
  • A unified graph-based exploration planner uses information gain, collision avoidance, and traversability constraints for legged and aerial robots.It also supports automated homing and travel to previously detected frontiers.Source: §1 · p. 3
  • Customized 5.8 GHz WiFi, ANYmal-deployed communication breadcrumbs, and Armadillo’s high-gain antenna provide bidirectional links to the Base Station.The links support supervisor control and reporting detected objects of interest.Source: §1 · p. 3

2 Related Work

Related work spans underground exploration, localization, planning, and multi-robot operation. CERBERUS situates its system-level effort within this developing research community.

2 Related Work

  • Earlier underground exploration work includes topological edge and intersection methods, aerial contour and junction detection, and open-space attraction approaches (Silver et al., 2006; Baker et al., 2004; Rogers III et al., 2017; Mansouri et al., 2019a; Mansouri et al., 2019b; Kanellakis et al., 2019).Groundhog demonstrated pioneering exploration capacity, while Jacobson et al. (2020) focused on subterranean localization challenges.Source: §2 · p. 4
  • The review frames subterranean autonomy as a broad challenge spanning perception, planning, control, communications, and mobility, with precedents in MAGIC and underground robotics.The cited work also includes systems overviews and rover navigation solutions (Morris et al., 2006; Nov´ak et al., 2015; Maity et al., 2013; Tardioli et al., 2019; Bakambu and Polotski, 2007; L¨osch et al., 2018).Source: §2 · p. 4
  • After the SubT Challenge began, work addressed multi-sensor mapping, legged-robot search, multi-robot exploration, visual interest, and robust localization.Other approaches included a hexapod design, virtual surfaces for ground mapping and planning, and a blimp for long-endurance missions.Source: §2 · pp. 4–5

3 CERBERUS System-of-Systems

The SubT Challenge targets autonomous exploration, mapping, and artifact search in underground settings with difficult terrain, visibility, communications, and access. CERBERUS responds with a co-deployed legged-and-aerial system whose robots share maps and report artifact positions in DARPA’s global frame.

3 CERBERUS System-of-Systems

  • Underground robots face dynamic rough terrain, limited visibility, poor communications, and difficult access, motivating autonomy research across several fields.The challenge incentivizes teams to explore, navigate, map, and search these environments.Source: §3 · p. 5
  • The Systems track includes physical Tunnel, Urban, and Cave scenarios, while the Virtual track evaluates software-only solutions in simulation.The Cave Circuit was cancelled; the Final Event was planned to combine prior circuit challenges.Source: §3 · p. 5
  • Teams search for 20 initially unknown artifacts, scoring one point when a reported position falls within 5 m of its surveyed location in DARPA’s coordinate frame.Artifact classes vary by circuit, with Survivor, Cell Phone, and Backpack common to all events.Source: §3 · p. 5
  • Tunnel and Urban courses each had two scored runs, with 30 minutes for setup and 60 minutes to explore and report artifacts per run.Circuit scores summed each team’s best run on each course.Source: §3 · p. 5
  • CERBERUS combines legged robots for complex terrain with aerial robots, especially collision-tolerant designs, for navigation less constrained by terrain.The vision combines these platforms with multimodal and multi-robot localization and mapping, artifact detection, exploration planning, and robot-deployable communications.Source: §3 · pp. 5–6
  • A single Human Supervisor oversees quadrupeds and aerial robots; ANYmal deploys communication nodes, and the robots collaborate on a consistent map.Flying systems, particularly collision-tolerant robots, can scout or detect artifacts in chimneys and shafts.Source: Figure 2 · p. 6

4 CERBERUS Robots

CERBERUS combines modified ANYmal ground robots, aerial scouts, and the Armadillo rover for underground exploration. The section describes platform-specific hardware and software adaptations for terrain, autonomy, and communication.

4.1 Subterranean Walking Robots - ANYmal B SubT

  • ANYmal B SubT adapts ANYmal B300 quadrupeds for underground deployment, leveraging robust, agile locomotion and 1–2-hour endurance.The platform uses three compliant joints per leg and can step over or crawl under obstacles.Source: §4.1 · p. 6

Hardware Modifications

  • For Urban Circuit terrain mapping, the Robosense BPearl was selected after the Pico Monstar and RealSense proved prone to reflections and environmental conditions.The BPearl was chosen for its accuracy, robustness across conditions, and 90° field of view; an auxiliary GPU supports parallel terrain mapping and neural-network object detection.Source: Hardware Modifications · p. 7

Software Modifications

  • After corrupted terrain representations disrupted Tunnel Circuit locomotion, Urban Circuit ANYmal used a reinforcement-learning controller relying on proprioception rather than a local terrain map.A state machine coordinates exploration planning and path tracking, with safety responses to obstacles and non-traversable terrain.Source: Software Modifications · pp. 8–10

4.2 Subterranean Aerial Robots

  • Conventional Matrice 100 scouts and collision-tolerant Gagarin flyers complement ground robots by reaching areas inaccessible to them.Charlie adds thermal vision for localization in visually degraded settings and thermal object detection; aerial scouts have limited flight endurance.Source: §4.2 · pp. 10–12

4.3 Roving Robots

  • Armadillo complements the walking and flying robots by carrying a high-gain antenna and providing wired connectivity to the Base Station.It carries a 300 m optical-fiber reel and was manually controlled by the Human Supervisor.Source: §4.3 · pp. 12–13

5 Exploration Autonomy

CERBERUS uses GBPlanner across legged and aerial robots to explore an incrementally updated 3D occupancy map, balancing local exploration with global frontier travel and homing. Its planning accounts for obstacles, geofences, endurance, and legged-robot traversability.

5 Exploration Autonomy

  • GBPlanner plans over a fixed-resolution 3D occupancy map, distinguishing free, occupied, unknown, and geofenced regions while excluding residual unknown space that robot constraints make infeasible to explore.The deployed voxel edge length was 0.2 m.Source: §5 · p. 13

5.1 Multi-Level Local Exploration Path Planning

  • Local planning samples collision-free paths in a graph around the robot and scores expected newly observed volume using the depth sensor’s range and field of view.Dijkstra’s algorithm supplies paths from the current configuration to candidate destinations.Source: §5 · pp. 13–14, §5.1 · p. 14, Figure 11 · p. 14
  • In multi-level environments, denser vertical graph sampling and an exploration reward for changing levels encourage routes through stairs or other inter-storey connections.A Human Supervisor selected the local exploration mode in the deployed implementation.Source: §5.1 · p. 15
  • Legged paths are checked for traversability during execution; encountering unknown non-traversable terrain stops the robot, marks a geofence, and triggers replanning, which can slow exploration.The planner assumes areas beyond the smaller elevation map are traversable; longer-range elevation mapping would be needed to address this limitation.Source: §5.1 · p. 15

5.2 Global Exploration Path Planning

  • After local completion, global planning selects frontier routes using expected exploration gain and remaining endurance, while maintaining a return-to-home path for the time limit or battery boundary.The global graph prunes frontiers that no longer offer high gain and groups similar candidate paths before retaining principal paths.Source: §5.2 · pp. 15–16
  • Local and global paths are shifted farther from obstacles to produce safer routes before robot-specific planners and controllers execute them.Source: §5.2 · p. 16

6 Multi-Modal Perception for Multi-Robot Mapping

CERBERUS combines complementary onboard sensing and centralized multi-robot mapping for subterranean localization and situational awareness. Individual robots can explore without a Base Station connection; when maps arrive, the server optimizes a shared map and artifact locations.

6 Multi-Modal Perception for Multi-Robot Mapping

  • Underground darkness, low texture, geometric self-similarity, thermal-flatness, and obscurants make reliance on one sensing modality unsuitable; CERBERUS combines visual and thermal cameras, LiDAR, and IMUs.Each robot maintains an onboard map, while the Base Station mapping server combines maps into a globally optimized view.Source: §6 · pp. 16–17, Figure 12 · p. 17

6.1 Time Synchronization and Calibration

  • Small timestamp inconsistencies can create significant systematic mapping errors when distant points are observed, because small rotations translate into large positional changes at range.The robots used several synchronization and calibration approaches, including exposure compensation and camera-to-sensor calibration.Source: §6.1 · pp. 17–18

6.2 Complementary Multi-Modal Localization and Mapping

  • CompSLAM hierarchically fuses visual and thermal imagery, LiDAR depth, and inertial cues, using health checks to pass trustworthy motion estimates into LiDAR odometry and mapping.The VTIO estimator uses an EKF; scan alignment and submap refinement are independently checked before pose estimates propagate.Source: §6.2 · p. 18, Figure 13 · p. 19

6.3 Global Multi-Robot Mapping

  • Robots build pose-graph submaps onboard and transmit them to the Base Station, reducing required robot-to-server bandwidth while using modest onboard computing.Submaps contain visual landmarks and LiDAR data; transfers can resume after connection losses.Source: §6.3 · p. 19
  • The mapping server connects robot maps through shared visual landmarks or LiDAR alignments, incorporates DARPA-frame alignment, and optimizes a globally consistent multi-robot map.The illustrated graphs show robots merged through either the DARPA gate or shared landmarks.Source: §6.3 · pp. 20–21, Figure 14 · p. 20
  • The optimized global map gives the supervisor a combined view and supports improved artifact locations in the DARPA frame; onboard-only artifact locations remain available when the global map is not.Source: Figure 12 · p. 17, §6.3 · p. 21

7 Artifacts

CERBERUS detects artifacts onboard, filters and attaches their reports to robot maps, and uses the globally optimized map to refine their DARPA-frame locations. The system also supports supervisor review and manual artifact input.

7 Artifacts

  • Competition scoring awards a point for each report that correctly identifies an artifact class and places it within 5 m of its ground-truth location.The artifact types varied between the Tunnel and Urban circuits.Source: §7 · p. 21

7.1 Detection and Classification

  • YOLOv3 detects visible artifacts from training images collected across robot platforms, laboratory backgrounds, and subterranean lighting conditions.Training sets ranged from 1,915 cell-phone images to 3,519 fire-extinguisher images.Source: §7.1 · p. 21, Table 3 · p. 21
  • The supervisor can inspect image streams and manually add, relabel, move, or remove map artifacts; Bluetooth detections support cell-phone reporting, and ground robots used CO2 sensors for gas detection.Source: §7.1 · pp. 21–22

7.2 Map Attachment

  • Onboard filtering casts detection-box rays into the map, uses the median point as an initial position, and combines nearby detections with class-specific Bayesian filters until a confidence threshold confirms an artifact.Confirmed detections are summarized for the supervisor, and their spheres are frozen to prevent duplicate reports.Source: §7.2 · p. 22, Figure 15 · p. 22
  • Artifact reports attach each filtered robot-frame position to its timestamped pose-graph node, allowing the global map to compute an optimized DARPA-frame position.For Bluetooth phone detections and Urban Circuit gas artifacts, the supervisor could help locate the object on the map or image.Source: §7.2 · pp. 22–23

8 Networking

CERBERUS built a robot-extended 5.8GHz network for underground communications and paired it with software for multi-robot ROS data exchange over unreliable links.

8 Networking

  • The 5.8GHz network carries mapping and artifact data; ground robots and dropped nodes form its mesh backbone, while aerial scouts join as clients.Ground nodes connect dynamically to route data; aerial robots connect to the nearest access point because they lack mesh-capable devices.Source: §8 · p. 23, §8.1 · p. 23

8.2 Base Station Network

  • The Base Station used a directional antenna, while Armadillo’s 300m optical fiber, pole antenna, and thrown WiFi node extended connectivity into the circuits.The pole and thrown node served stairways; Armadillo also extended WiFi around nearby corners and corridors.Source: §8.2 · p. 24

8.3 Deployable Mesh Nodes

  • ANYmal carried four battery-powered mesh nodes, with roughly 100m ideal straight-tunnel range and observed bandwidth near the first node of 200Mbps.Range varied with obstructions, multipath, and reflections, so the Human Supervisor chose deployment locations using ping quality and surrounding geometry.Source: §8.3 · p. 24

8.4 Nimbro - Multi-Master Network

  • Nimbro Network decoupled robot rosmasters and selectively transmitted ROS data over unreliable links; chrony synchronized robot clocks to the Base Station.The team chose explicit stream selection over automatic data decisions, given more than 800 topics and 400 services on an ANYmal and the risk of unnecessary traffic.Source: §8.4 · p. 25

9 Operator Interface

CERBERUS adapted operator control around the single-supervisor constraint, moving from per-robot PCs toward shared control interfaces while providing centralized artifact review and fallback reporting.

9 Operator Interface

  • Between circuits, CERBERUS improved multi-agent GUIs and reduced exchanged data to better meet the single-supervisor requirement.The requirement limits coordination and monitoring to one person at the Base Station, with the aim of minimizing human intervention during autonomous missions.Source: §9 · p. 25

9.1 Robot Control

  • Tunnel control burdened the supervisor with multiple Operator PCs; Urban Circuit consolidated robot control into two PCs and sent high-bandwidth sensor data only on request.Robots still had dedicated PCs for setup before deployment, while Nimbro Network exported only selected ROS topics and services to the interfaces.Source: §9.1 · p. 26

9.2 Artifact Reporting

  • The ARI combined global-map artifact review, video, automatic detections, and manual reporting, with annotated image context and validity feedback for submitted reports.Supervisors could accept or modify detections, add missed artifacts, and resubmit initially rejected reports after map improvements.Source: §9.2 · pp. 26–27

9.3 Fallback Solutions

  • When optimized maps were unreliable or stale, the supervisor could report artifacts and maps using unoptimized poses; centralized state was also backed up for safe restarts.Packages could be relaunched after crashes, and artifact positions received at the Base Station were periodically saved to disk.Source: §9.3 · pp. 27–28

10 Experimental Evaluation

The evaluation follows CERBERUS from pre-event field integration through Tunnel and Urban Circuit deployments, then assesses performance and operational limits. The team placed sixth in the Tunnel Circuit and fifth in the Urban Circuit, with Urban deployments showing gains in mapping and autonomous operation alongside persistent autonomy and communication challenges.

10.1 Field Tests prior to the Circuit events

  • Field tests across varied underground and urban sites supported integration of heterogeneous research prototypes, and component unification improved interoperability for the Urban deployment.Only a subset of sites hosted full-team tests; STIX offered qualified teams a representative environment under competition rules.Source: §10.1 · p. 28

10.2 DARPA Subterranean Challenge – Tunnel Circuit

  • Tunnel deployments reported six artifacts and five points; the Alpha Aerial Scout traversed 180 m autonomously, while Armadillo reported four artifacts after being repurposed.The Alpha’s exploration was limited by a bounding box set incorrectly by the Human Supervisor; Armadillo’s odometry used LiDAR alone and drifted, though its final APE remained within the ±5 m reporting margin.Source: §10.2 · pp. 29–30

10.3 DARPA Subterranean Challenge – Urban Circuit

  • Urban deployments reported eleven artifacts, scored seven points, and placed fifth of ten; global mapping lowered APE versus onboard estimates while keeping APE within the ±5 m reporting margin.The exploration planner also produced paths through narrow openings and refined paths for obstacle clearance in the Beta course.Source: §10.3 · pp. 30–31, Figure 22 · p. 33, Table 5 · p. 34

10.4 Circuits Evaluation

  • CERBERUS finished sixth of eleven teams in the Tunnel Circuit and fifth of ten in the Urban Circuit.The circuits also exposed limitations that informed the team’s lessons learned and future work.Source: §10.4 · p. 34

10.4.1 Tunnel Circuit

  • Tunnel locomotion failures included corrupted elevation maps and wheel slippage; Alpha demonstrated autonomous exploration, while Armadillo’s last-run role yielded four additional points.The elevation-map issues were attributed to foot slippage or corrupted RealSense depth data; the team later changed sensing and locomotion components.Source: §10.4.1 · p. 35

10.4.2 Urban Circuit

  • In the Urban Circuit, quadrupeds were deployed reliably and increased speed from 0.2 to 0.45 m/s, but traversability and state-machine issues slowed missions; Alpha returned autonomously and Gagarin descended stairs collision-tolerantly.The shared exploration planner worked across deployed robots with robot-specific parameter changes, but lacked shared exploration information and direction biasing.Source: §10.4.2 · pp. 36–37
  • Urban multi-robot mapping improved pose estimates, but sensor differences and longer global-map optimization limited obscurant suitability and map freshness for the supervisor.Robots sent 4–7 MB submaps every 30 s; combined with teleoperation and fallback information, bandwidth could reach 6–9 MBps, contingent on deploying enough WiFi nodes.Source: §10.4.2 · p. 37
  • Communications benefited from the tethered rover and WiFi nodes, but breadcrumb placement and orientation, WiFi quality, and tether handling remained practical challenges.The fiber link helped in urban environments with thick walls and multiple turns, although teleoperating the rover demanded supervisor attention to avoid cable tangling.Source: §10.4.2 · pp. 37–38

11 Main Lessons Learned and Discussion

CERBERUS’s field experience showed improved artifact-location estimates alongside persistent challenges in perception, planning, autonomy, communications, and system integration. The team highlights relevant field testing and simulation as important to developing resilient subterranean robotics.

Perception and mapping

  • Heterogeneous sensors complicate shared localization and mapping, while limited CompSLAM–M3RM integration left scope for sharing onboard features and maps.The team notes that camera, lens, and illumination differences affect loop closure and map merging; bidirectional map sharing could support collaborative planning and aerial launches deeper underground.Source: §11 · p. 39

Artifact reporting

  • Centralized M3RM map optimization reduced individual-robot APE and improved map and artifact position estimates despite onboard drift.Artifact reports score only when reported positions are within 5 m of true positions; inconsistent illumination caused missed detections, and reporting Bluetooth presence for supervisor localization yielded better cell-phone results.Source: §11 · p. 39

Exploration planning

  • GBPlanner’s depth-sensor volume gain does not ensure artifact detection, and short-range traversability maps prevent secure long-term legged planning.The team considers receding-horizon planning; artifact search may need to account for multiple sensor modalities and objectives.Source: §11 · p. 40

State machine

  • Urban Circuit legged robots sometimes could not recover from unforeseen obstacles, particularly outside WiFi range, after state-machine corner cases went unhandled.The team is considering behavior-based autonomy and a watchdog for recovery from partial software failures.Source: §11 · p. 40

Locomotion

  • Perception-limited model-based control gave way to resilient blind neural locomotion, while wheel slippage prompted hybrid-gait MPC and standardized point-foot deployment.The new wheel controller dynamically selects a hybrid of powered rolling and legged stepping (Bjelonic et al., 2021).Source: §11 · pp. 40–41

Communications and supervision

  • Reliable communications remained challenging; the rover extended WiFi but consumed supervisor time, while multi-robot supervision lacked a unified, easy-to-use interface.The proposed rover practice was to drive it through the first main turns and park it; Nimbro Network had a positive performance record.Source: §11 · p. 41

Integration and simulation

  • Geographic separation and research-prototype systems made integration difficult, while increased simulation use supported investigation of more complex exploration and multi-robot behaviors.Software was later unified across more robots, and the team began conducting missions with CERBERUS robots together in Switzerland.Source: §11 · pp. 40–41

Development practice

  • Field testing in relevant settings guided autonomy research, while five-minute logging chunks, weekly shakeouts, and open-source releases supported development and community use.Weekly shakeouts helped find software defects and standardize setup but did not replace underground field deployments; open-sourcing was intended to support reproducibility and comparison.Source: §11 · p. 42

12 Priorities for Future Work

CERBERUS’s future-work priorities build on deployment lessons through improved platforms, perception, mapping, exploration, communications, and field testing.

Legged and flying robots

  • ANYmal C is intended to replace ANYmal B, while high-speed collision-tolerant aerial robots are to be deployed where ground robots cannot reach.The new ANYmal C integrates a learning-based locomotion controller and the CERBERUS multi-modal sensor and processing suite.Source: §12 · p. 43

Resilient perception

  • Planned perception improvements include robot-specific odometry, degradation detection, and tighter multi-modal sensor integration for localization and mapping.Leg odometry is one example of an external estimate the team aims to incorporate.Source: §12 · p. 43

Multi-robot mapping

  • Bidirectional sharing of maps and explored-area knowledge is intended to improve collaborative localization, mapping, planning, and breadcrumb deployment.Shared maps could also support aerial launches deeper into the circuit and artifact reporting without independent initial DARPA-frame alignment.Source: §12 · p. 43

Autonomous exploration and visual search

  • Future exploration will co-optimize volumetric gain and artifact coverage while accounting for each platform and incorporating recovery behaviors.The stated goal is autonomous exploration and artifact search by each robot, with human input reserved for high-level decisions.Source: §12 · p. 43

Communications and field testing

  • CERBERUS is developing a longer-range, more reliable WiFi baseline and improved ANYmal C breadcrumbing, alongside more frequent large-scale field testing.The team aims for monthly tests with a character resembling the challenge circuits; the Armadillo rover will deploy a high-gain antenna while maintaining an optical-fiber connection to the ground.Source: §12 · p. 43

13 Conclusion

The paper reports CERBERUS’s technological progress through the DARPA Subterranean Challenge Urban Circuit. It describes a legged, aerial, and roving robot system, its autonomy and communications, and its circuit performance and lessons.

System and scope

  • CERBERUS combines legged and flying robots, including collision-tolerant designs, with a roving platform and autonomy, artifact-detection, and communication systems.The paper reports deployments in both Tunnel and Urban Circuits, along with lessons learned and planned future work.Source: §13 · pp. 43–44

Reported contribution

  • The reported research facilitated some resilience in CERBERUS robots operating in visually degraded, large-scale, narrow, and complex underground settings.The work also reports the team’s performance in both circuits and discusses critical lessons from those experiences.Source: §13 · p. 44

Community contribution

  • The paper presents open-source code and future activities as contributions toward community efforts in resilient robotic autonomy underground.Its stated scope includes a brief overview of work planned before the final Subterranean Challenge event.Source: §13 · p. 44

A List of Open Source Packages

The appendix presents a selected, non-exhaustive list of CERBERUS open-source packages spanning perception, mapping, planning, calibration, and navigation.

Exploration planning

Calibration and detection

Odometry and mapping

  • ROVIO provides visual-inertial odometry; Maplab includes M3RM multi-robot mapping, while Voxblox and Voxgraph support voxel-based and globally consistent volumetric mapping.Source: A List of Open Source Packages · pp. 44–45

Terrain mapping

Simulation Models

CERBERUS released robot models for the official SubT Virtual Track Challenge simulation environment. The release includes models of its ANYmal B and two aerial scouts.

Simulation Models

  • The SubT Challenge Systems and Virtual Competitions are interconnected through robot models and potential autonomy benefits.Systems teams deliver robot models for the Virtual Competition, while Virtual Competition progress can benefit Systems Competition autonomy.Source: Simulation Models · p. 45
  • CERBERUS robot models were released for use in the official SubT Virtual Track Challenge simulation environment.The paper notes that Virtual Competition progress can benefit autonomy in the Systems Competition.Source: Simulation Models · p. 45
  • Models released so far include the ANYmal B “SubT,” Alpha Aerial Scout, and Gagarin Aerial Scout.The paper says models of most robots fielded in the Tunnel and Urban Circuits have already been released.Source: Simulation Models · p. 45

B Videos from the Tunnel and Urban Circuits

The paper assembles videos highlighting CERBERUS deployments in the Tunnel and Urban Circuit events. The videos show deployments on multiple Urban Circuit courses and an aerial scout in a Tunnel Circuit course.

B Videos from the Tunnel and Urban Circuits

Related papers

Loading 2201.07067v1…