Overview
The DARPA Subterranean (SubT) Challenge aimed to advance practical robotics for extreme underground environments. In September 2021, an extensive limestone cave network near Louisville, Kentucky hosted the event's final. The cave was dark, damp, and dusty, and teams gathered to tackle a complex underground course designed by the Defense Advanced Research Projects Agency (DARPA) as the culmination of the three-year SubT competition.
The SubT Challenge Format
DARPA launched the SubT Challenge in early 2018 to accelerate capabilities for robots operating in three underground domains: man-made tunnels, subterranean urban infrastructure, and natural caves. To score well, robotic teams had to cooperate, traverse unknown terrain for kilometers within strict time limits, build maps of the environment, search for simulated artifacts, and report their locations accurately. To make scenarios realistic for potential emergency-response use, robots had to operate amid darkness, dust, smoke, and deliberate collapses controlled by DARPA.
The competition offered direct funding and prize money totaling several million dollars and encouraged collaboration among leading academic institutions and industry teams worldwide. Each team could experience all three environments in separate circuit events.
Preliminary Rounds and the Finals
In August 2019, the tunnel circuit took place at the U.S. National Institute for Occupational Safety and Health experimental coal mine near Pittsburgh, where many teams lost contact with robots within the first turn. Six months later, the urban circuit was held at an unfinished nuclear plant in Satsop, Washington, and teams experimented with a range of communications strategies, from tethered Ethernet cables to battery-powered mesh nodes that robots could drop like breadcrumbs to extend connectivity. The cave circuit was scheduled for fall 2020 but was canceled due to the COVID-19 pandemic.
In the Louisville final, the emphasis shifted from communications to autonomy. As in the preliminaries, people were not allowed on the course, and each team was permitted only one remote operator to interact with their robots, making direct teleoperation impractical. The most viable approach was autonomous decision-making by robotic teams about where to go and how to get there.
DARPA built a roughly 1 km course inside an existing cavern for the final. Staff connected shipping containers into complex networks, many carefully dressed to resemble mining tunnels and natural caves. Areas representing offices, warehouses, and even a subway station were constructed from scratch for the urban portion of the course. Teams had one hour to locate as many as 40 simulated artifacts and report their positions back to the base at the course entrance, a difficult task because direct communications from deep in the course were not possible.
Results and Mapping Performance
Eight teams reached the final, most bringing coordinated fleets of robots. Wheeled platforms offered reliable mobility, four-legged robots showed surprising capability on complex terrain, and drones provided broad reconnaissance for large caverns.
By the end of the final, two teams each reported locating 23 artifacts: the Cerberus team (a collaboration including University of Nevada, Reno; ETH Zurich; Norwegian University of Science and Technology; UC Berkeley; Oxford Robotics Institute; Flyability; and Sierra Nevada Corporation researchers) and the CSIRO Data61 team (a collaboration of CSIRO's Data61, Emesent, and researchers from Georgia Tech). The tie-breaker was which team found the last object faster; Cerberus finished 46 seconds ahead and took first place.
Although second in ranking, the CSIRO team produced an impressive result: their map of the course differed from DARPA's ground-truth map by less than 1%, effectively matching a map made by a human expert team over many days. That level of mapping accuracy is the kind of concrete progress DARPA sought, according to the program manager Tim Chung.
Why SubT Matters for Robotics
Tim Chung commented that underground environments expose many problems often considered separately in other contexts. The sheer scale of underground infrastructure presents abundant opportunities to improve perception, understanding, and navigation capabilities, as well as challenges in engineering integration and core design. Addressing these issues could change robotics over the next 5, 10, or 15 years.
Autonomy and the Human Role
Early rounds of the competition showed many teams relying on close operator guidance with low-level commands. Teams quickly realized they needed higher levels of autonomy. Complete autonomy is difficult to achieve in practice, and humans will continue to play an important role, though that role will evolve. Operators are likely to progress from direct controllers to supervisors who set task-level goals and intervene when needed. In the final, robots explored and searched while supervisors focused on higher-level tasks, demonstrating a shift from "remote control" to "robots handle routine tasks, humans handle higher-level decisions."
Key Technical Challenges in Underground Environments
Navigation and reasoning about traversability remain open problems. Robots that better understand where they are and where they are looping can traverse environments faster. While mapping performance can be one to two orders of magnitude faster than humans, robot traversal speeds remain relatively slow. Improving speed by an order of magnitude would significantly affect emergency-response scenarios where every minute counts.
Technical and Commercial Impact
Many technologies demonstrated in the SubT Challenge are moving toward productization and commercialization, shortening the time before robotic tools reach emergency responders. DARPA's approach encouraged multiple teams to develop robust, deployable systems rather than only novel research prototypes. This has accelerated the readiness of field-deployable robotic solutions.
Interview Excerpts: Team Perspectives
Why underground is both a challenge and an opportunity
Navinda Kotagede (CSIRO Data61): Being underground means no GPS, unreliable communications, and difficult mobility. These conditions also appear in other contexts, such as the response to the Fukushima nuclear accident or dense environments like the Amazon rainforest. Component technologies matured for SubT will apply to many domains beyond caves and tunnels.
Human involvement in human-robot teams
Kotagede: Human roles lie between two extremes: pressing a button and fully human-in-the-loop remote control. If humans are completely out of the loop, systems can stall when communications are fragile. The middle ground is "human on the loop," where a supervisor sets task-level goals and the team continues operating if contact is lost. Humans provide valuable scene-level reasoning, which they are much better at than machines.
What SubT forced teams to achieve
Kotagede: Successful field deployment requires many technologies to work reliably together. Three years ago we had interesting components but not a system that could run unattended for long. The competition forced us to build robust, reliable systems that could be deployed in real scenarios. Once a reliable autonomous team exists, new tasks and applications become possible.
How the team produced high-quality maps
Kotagede: The official ground-truth map took a specialist company about 100 labor hours with expensive equipment. Our fleet produced a comparable map in under an hour because the mapping software represents about 15 years of research and iterative testing in harsh conditions. Turning theoretical solutions into robust real-world systems required pushing mapping systems to their limits over the past decade.
Why participate in SubT
Kostas Alexis (Cerberus): The competition motivated teams to build systems that work reliably outside the lab. The combination of societal impact and technical challenge was compelling. In a competitive setting you cannot just work in the lab or publish papers; you must deliver systems that work under continuous, difficult conditions.
Major challenges for teams
Alexis: Communication cannot be relied on, terrain is complex, and the course scale is large. Robots must be both autonomous and robust at the system level. DARPA understood these capabilities were not present at the outset but committed to a three-year program to accelerate progress, which proved effective in advancing the state of the art.
Future human roles as autonomy increases
Alexis: Robots now produce excellent maps and perform object detection, but scene-level reasoning and potential interactions with objects remain weak. If time is unconstrained, full automation could achieve comprehensive coverage. When time is limited and maximal exploration is needed, human reasoning over the data remains valuable. Social factors also matter: humans ultimately make final decisions for critical operations.
Legged robots and combined approaches
Alexis: The competition demonstrated that legged robot systems can be deployed in challenging conditions. Combining legged platforms with drones can be especially effective for complex ground and underground scenarios.
Timeline to practical applications
Alexis: Commercialization driven by SubT is likely to be faster than typical academic research cycles. Timeframes could be months to about a year for many industrial use cases. For disaster response, achieving near-100% reliability and handling certification and assurance across use cases are significant issues. Those requirements will take longer. It is likely these systems will appear first in industrial applications before disaster-response deployment.
