AI Robot Swarms in the DARPA Subterranean Challenge

The DARPA Subterranean Challenge pushed the boundaries of robotics by tasking heterogeneous teams of autonomous robots with exploring and mapping pitch-black caves, human-made tunnels, and ruined mines. Operating entirely without GPS and in environments that quickly severed communications with human operators, these robotic swarms relied on decentralized multi-agent artificial intelligence to survive and succeed. This article breaks down the core technologies behind their success, specifically focusing on Simultaneous Localization and Mapping (SLAM), peer-to-peer mesh networks, autonomous task allocation, and multi-modal sensor fusion.

Because satellite signals cannot penetrate deep underground, robots could not use GPS to know their locations. Instead, teams utilized advanced Simultaneous Localization and Mapping (SLAM). Each robot built a 3D map of its immediate surroundings in real-time while tracking its own movement within that map.

To do this in pitch-black, dust-filled, and smoke-obscured environments, robots relied on multi-modal sensor fusion. They combined data from LiDAR (laser scanners), thermal cameras, Inertial Measurement Units (IMUs), and proprioceptive sensors (which measure internal states like wheel slip or joint angles in legged robots). By fusing these inputs, the AI could accurately map a collapsing tunnel even if a sensor failed or became blinded by thick dust.

Decentralized Multi-Agent Coordination

Relying on a central computer or a human operator to direct the robots was impossible because rock walls rapidly blocked radio signals. If a robot waited for human instructions after losing connection, it would become useless. The solution was decentralized multi-agent AI.

Every robot in the swarm operated with its own onboard brain, making independent decisions based on its immediate environment. Instead of a central hive mind, the robots acted like a team of independent explorers. They used frontier-based exploration algorithms to identify the boundaries between mapped and unmapped areas. When two or more robots were within communication range of each other, they automatically negotiated to divide the workload. They shared their current trajectories and assigned themselves to different unexplored frontiers to prevent duplicating efforts.

Overcoming the Communication Blackout

To get data back to the human operators stationed outside the cave entrance, the robots had to build their own communication infrastructure on the fly. The teams used wheeled and legged robots—such as Boston Dynamics' Spot or custom-built rovers—to carry and physically drop wireless communication nodes onto the ground as they ventured deeper.

These breadcrumbs formed a dynamic, peer-to-peer mesh network. However, nodes were limited, and robots eventually had to push beyond the network's reach into total isolation. In these "comms-denied" zones, a robot would continue exploring, mapping, and searching for targets (like survivors, backpacks, or gas leaks) completely on its own. Once its local memory was full or its battery ran low, the AI directed the robot to backtrack until it re-established a link with the mesh network or encountered another robot.

Collaborative Map Merging

When isolated robots finally reconnected with each other or the mesh network, they needed to combine their individual maps into one unified picture. Sharing massive, raw 3D point clouds over weak underground radio links was impossible due to severe bandwidth limits.

To solve this, the onboard AI compressed the maps into lightweight "pose graphs"—simplified mathematical representations of the robot's path and key environmental features. When robots met in the dark, they exchanged these compact graphs. The AI used "loop closure" algorithms to recognize common physical landmarks that multiple robots had seen independently. By locking onto these shared landmarks, the system stitched the individual maps together, creating a single, highly accurate 3D map of the entire subterranean network for the human team waiting at the surface.