CERBERUS in the DARPA Subterranean Challenge

In short: underground spaces are dangerous and unknown, so DARPA asked teams to build robots that could scout them for first responders, and turned it into a timed treasure hunt.

Underground settings range from tunnels several kilometres long to multi-level urban structures, from narrow cave passages and vertical shafts to vast chambers. They can be hard and dangerous for people to enter, which is why robots that give a small team of operators rapid situational awareness before anyone goes in are valuable for both military and civilian first responders.

The DARPA Subterranean (SubT) Challenge was a three-year, $82 million competition that brought together more than 300 competitors from 20 teams in 11 countries. It had a Systems track (real robots in real underground courses) and a Virtual track (software only, in simulation). The Systems track visited a mine tunnel network (Tunnel Circuit, August 2019) and multi-level urban underground structures (Urban Circuit, February 2020); the Cave Circuit (August 2020) was cancelled. The Final Event in September 2021 combined all three.

The rules that shape everything

• Each team had 60 minutes to find as many as possible of 40 artifacts: mannequin “survivors”, cellphones, backpacks, drills, fire extinguishers, vents, gas-filled rooms, helmets, ropes and a custom “SubT cube”.
• A report scored a point only if it named the right type and placed it within 5 m of its true position. The number of reports was limited, to discourage guessing.
• Only one team member, the Human Supervisor, was allowed to manage and talk to the deployed robots.

With one human and a radio link that fades with depth, most of the work has to be done by the robots themselves. That is the central demand of the contest, and of this paper.

In short: DARPA named four areas where robots break down below ground, and a winning team needs all four at once.

• Autonomy. Map, navigate and search complex, changing places “without substantial human interventions”.
• Perception. Keep working in dust, fog, mist, water, smoke and darkness, with obscured views.
• Networking. Get information back to the Human Supervisor despite limited line of sight, varying rock and the way radio waves behave underground.
• Mobility. Cross narrow passages, sharp turns, big drops and climbs, slopes, steps, falling debris, mud, sand and water.

By the start of the challenge, teams were trying every kind of platform: wheeled, tracked, legged and flying robots. Decades of work on mapping and path planning existed, but the authors put it bluntly: underground environments “proved to be the domain where autonomy had many reasons to break”.

In short: CERBERUS was a “system of systems”: walking robots that explore, flying robots that reach where legs can’t, and a radio network the robots lay down themselves, all run by one human.

The team was an international partnership: the University of Nevada, Reno, ETH Zurich, NTNU, UC Berkeley, the University of Oxford, Flyability and Sierra Nevada Corporation.

What this means: the legged robots did the heavy exploring; everything else existed to give them reach (drones) and a voice back to the base (rover and relays).

A radio network the robots build themselves

Underground, radio propagation is hard to predict and depends heavily on the tunnel geometry. CERBERUS used an ad-hoc wireless mesh network at 5.8 GHz, in which the legged robots and the rover were themselves nodes. Each Carrier ANYmal ferried four communication-extender modules on its back and released them on the supervisor’s command by lowering its torso and switching an electromagnetic release. Each module was a sealed, water- and dust-proof unit with a Rajant BreadCrumb DX2 radio and a battery for two hours. Drones joined the mesh as clients.

Who is in charge? Supervised autonomy

The team ran in supervised autonomy: the Human Supervisor read the maps arriving at the base station and gave high-level tasks; each robot carried them out on its own. The paper lists four ways to task a robot:

1. explore an unknown area while staying in radio range, backtracking to a connected spot if the link is lost;
2. explore beyond radio range for a set time budget, then come back into range;
3. walk to a chosen frontier near unmapped space, then start exploring on its own;
4. as a last resort, go to a waypoint in the map along a path the robot finds itself.

In short: a single exploration planner, shared by walking and flying robots, looks for the most informative path nearby and falls back on a mission-wide memory of unexplored openings.

At the core of CERBERUS’s autonomy was the graph-based exploration path planner, GBPlanner (Dang et al. 2020; Kulkarni et al. 2022). It works on a volumetric map of the environment, built from the robot’s sensor data and poses (Voxblox), and has two stages:

• Local exploration. Given the robot’s position, the current map and bounds on the exploration space, it samples a random graph inside a local box around the robot, then picks the path along which its range sensors would see the most unknown space. The planning is truly 3-D and, for ground robots, respects limits on slopes and other terrain features.
• Global mode. In complex, steep underground layouts the local stage can run out of useful paths. The planner then switches, automatically or when the supervisor asks, to a sparse graph built up over the whole known map. It stores the frontiers of exploration and the paths to them, so it can reposition the robot to unexplored areas, or return it home or to the last point with a radio link when its battery or time budget runs out.

Because the same planner drives legged and flying robots, the team could add or swap platforms without writing a new exploration strategy for each, which the authors list as one of their lessons.

In short: a legged-robot planner reshapes the exploration path to fit the ground, and a walking controller learned in simulation executes it, deciding for itself how much to trust its view of the terrain.

The final course had rails and hoses in the tunnel section, stairs in the urban section and a rough, slippery cave with gravel, rocks and a steep, narrow passage. Two layers turned an exploration path into safe footsteps:

• Navigation planner. A reachability-based planner for legged robots (Wellhausen & Hutter 2021) follows the exploration waypoints but can adjust the path locally. It uses an elevation map computed on the GPU (Miki et al. 2022) and a cost map of how traversable each patch is.
• Learned walking controller. A perceptive locomotion policy (Miki et al., Science Robotics 2022) tracks the velocity commands. It was trained entirely in simulation with reinforcement learning. To make it robust, training randomised external forces and torques on the body, the initial body and joint velocities, the total mass and the slipperiness of the ground.

The key property is how it combines its senses. The elevation map is essential for obstacles such as stairs, but dust, smoke and puddles add noise to the depth sensors. The controller was trained with a variety of noise in its terrain input and learned, end to end, how much to rely on each source: use the map when it looks reliable, for fast and smooth walking; ignore it when it looks degraded. The same training lets it cope when a sensor stops sending data altogether.

Earlier in the competition the team also tried different feet for ANYmal: point feet, flat feet and wheels. They settled on point feet as the most versatile across terrain types, avoiding the need for specialised designs. The drones used conventional cascaded position–attitude control, with fixed gains or model predictive control.

In short: every robot tracks its own position with a chain of sensor-specific estimators that skips any link that fails, and the base station then fuses all robots into one mission map.

Localisation and mapping (SLAM) rested on two ideas: sensors of different kinds fail in different situations, so combining them gives redundancy; and the software had to adapt to very different robots, from a 55 kg quadruped to a small drone. The result was a two-stage solution.

On each robot: CompSLAM

CompSLAM (Khattak et al. 2020) fuses visual and thermal images, LiDAR depth, inertial measurements and the legs’ kinematic estimates in a hierarchy. Each sensing modality (or group) has its own pose estimator. If its output passes data- and process-level health checks, it becomes the starting point for the next estimator in the hierarchy. If an estimator is judged inconsistent, the previous estimate is handed straight to the next one, skipping the failed step. So every robot always has an up-to-date pose and map.

At the base: M3RM, one map for the team

When in radio range, each robot incrementally sends its odometry and sensor data to the base station: visual-inertial factor graphs plus subsampled, compressed LiDAR scans, to fit the network’s limited bandwidth. The base station, with far more computing power, optimises them jointly into a globally consistent multi-robot map (M3RM, built on the maplab framework) and uses it to report optimised artifact positions. If that fails, the per-robot CompSLAM maps serve as a fallback; most drones reported only their onboard estimates. Splitting the job between robot and mission level avoids a single point of failure.

In short: scoring means naming the object and placing it within 5 m, so detection was paired with careful localisation, and a human confirmed every report.

All visual artifacts (everything except cellphones, gas and the SubT cube) were detected with YOLOv3, a fast image-detection neural network, trained on more than 40,000 labelled images collected underground. The team deliberately gathered images with obscurants, from different cameras, against varied backgrounds, under changing light and with motion blur. The network ran onboard every robot.

For each detection, the robot cast rays from the pixels inside the detection box into its volumetric map to get a first 3-D position, then refined it with later sightings as it moved. A binary probability filter, updated with every detection, suppressed false alarms.

Objects that cannot be seen were sensed differently: the cellphone and SubT cube by their Bluetooth signal strength, and gas with a CO₂ sensor. Within a window of measurements, the object was placed at the robot’s position where the signal was strongest. In every case the Human Supervisor inspected the detection and accepted or rejected it before a report was sent.

In short: CERBERUS tied for the most points and won on the tie-break, because its final report came earlier.

The Prize Round of the Final Event took place in the Louisville Mega Cavern, Kentucky, in September 2021, with eight Systems-track teams and nine Virtual-track teams. In the winning run CERBERUS deployed four ANYmal C SubT robots. Together they walked more than 1.7 km and correctly reported 23 of the 40 artifacts.

What this means: the top two teams were level, and the next best found 5 fewer artifacts. The tie-break rule, “the team that identified its last artifact earliest would win”, decided it: CSIRO Data61’s final point came with under 30 seconds left, CERBERUS’s with more than a minute left. CERBERUS won the challenge and the $2 million prize.

In short: legs plus wings, redundant perception and a human strategist worked, but robots still lack understanding, risk awareness and decentralised teamwork.

• Legged and flying robots are the right mix underground. CERBERUS relied almost entirely on walking and flying systems from the start, and by the Final almost all teams leaned heavily on them. Walking robots were the main explorers, with endurance and terrain ability beyond wheels or tracks; drones filled the gaps: vertical spaces and impassable ground. But small drones have short endurance, and the places that need them may be far from the entrance. The team worked on ANYmal carrying the RMF-Owl drone (“marsupial” deployment), finished only after the competition.
• Resilient perception and autonomy are the key technologies. Fusing sensors that fail in different ways (LiDAR does not need visual texture; cameras do) minimised breaking points. And one exploration policy that reflects the geometry of underground spaces could guide very different robots.
• Autonomy and human input work best together. Without prior infrastructure, underground spaces don’t allow reliable high-bandwidth links, so autonomy is what makes missions possible. When a link exists, the supervisor’s high-level commands improve where robots go. All the dense, moment-to-moment path decisions stayed with the robots.

What is still missing (the authors’ list)

• semantic understanding of the environment;
• comprehensive modelling of navigation risks, for example how perception conditions should influence path choices;
• tighter coupling between exploration planning and traversability;
• multi-robot SLAM that stays consistent under severe sensor degradation;
• efficient decentralised coordination of many robots.

The authors expect the same methods to transfer beyond the underground, to mapping construction sites, exploring new infrastructure, and inspecting ships and process tanks: places with the same mix of hard terrain, poor sensing and a need for autonomous mapping and object recognition.

About this paper: it is a short overview (a few pages, two figures, no experiments beyond the competition result). Numbers on this page are the ones it reports; the detailed evaluations live in the cited papers on GBPlanner, CompSLAM, M3RM, the navigation planner, elevation mapping and perceptive locomotion, and in the team’s Field Robotics paper on the Tunnel and Urban circuits.