TL;DR
Team CSIRO Data61 tackled the DARPA Subterranean Challenge by developing a heterogeneous fleet of ground and air robots with a novel approach: homogeneous sensing. This allowed for robust, decentralized multi-agent SLAM and exploration, enabling autonomous coordination even with intermittent communication, a critical factor for success in complex underground environments.
What Problem Does This Paper Address?
Exploring dangerous, unstructured subterranean environments poses significant challenges, primarily due to limited communication and the need for robots to operate autonomously. The DARPA Subterranean Challenge (SubT) aimed to push the boundaries of robotic autonomy in these scenarios, requiring teams to find artefacts in complex environments within a tight timeframe. Traditional approaches often rely on extensive communication infrastructure or highly specialized, expensive robots. This paper addresses the challenge of creating a coordinated, autonomous robotic team that can effectively explore and map such environments, even when communication is unreliable, by leveraging a heterogeneous fleet with a unifying sensing strategy.
Methods
The core of CSIRO Data61's approach was the homogeneous sensing strategy applied to a heterogeneous fleet of robots. This meant that despite using different robot platforms (tracked UGVs, legged UGVs, and UAVs), they all utilized a common sensor suite, primarily a spinning lidar and cameras, integrated into a modular "CatPack" for UGVs and the "Hovermap" payload for UAVs. This homogeneity was crucial for their decentralized multi-agent SLAM (Simultaneous Localization and Mapping) system, named "Wildcat." Each agent independently computed a global map, sharing SLAM frames peer-to-peer. This allowed for robust navigation and exploration even when disconnected from a central base station.
Key techniques included:
- Wildcat SLAM: A multi-agent SLAM system that computes global maps on each agent by integrating sensor data and sharing SLAM frames. This enabled a shared understanding of the environment across the fleet.
- 3D Frontier-based Exploration: Robots identified and pursued frontiers (boundaries between explored and unexplored space) to guide their exploration, a standard technique enhanced by the homogeneous sensing and shared map.
- Modular Perception (CatPack): A standardized perception module for UGVs that maximized robustness and enabled reuse of the perception and autonomy stack across different ground robots.
- Multi-Robot Task Allocation (MRTA): A market-based approach where agents bid on tasks based on estimated rewards, allowing for dynamic and decentralized coordination, especially when communication was lost.
- Marsupial UAV Deployment: Larger UGVs carried UAVs, allowing them to be launched in situ in areas inaccessible to ground robots, significantly expanding exploration capabilities.
Datasets
The paper mentions the use of a continuously extended annotated RGB dataset for training artefact detection models. This dataset contained over 13,000 images from various environments, including urban construction sites, natural caves, and underground areas, with labelled instances of fire extinguishers, cell phones, survivors, vents, helmets, ropes, backpacks, and drills. This dataset was augmented with images from UPenn, Shivakumar et al. (2019), MSCOCO (Lin et al., 2014), and ImageNet (Deng et al., 2009).
Key Insights and Ideas
- Homogeneous Sensing is Key to Heterogeneous Autonomy: The paper strongly advocates that unifying the sensing capabilities across a diverse robot fleet is a more effective path to robust multi-agent autonomy than focusing solely on platform heterogeneity. This simplification enabled a powerful, decentralized SLAM and coordination system.
- Decentralized SLAM for Communication Resilience: Relying on peer-to-peer SLAM frame sharing, rather than a centralized map, proved critical for maintaining situational awareness and coordination even with intermittent communication, a common issue in subterranean environments.
- Marsupial UAVs Unlock New Exploration: The strategy of having larger UGVs carry and deploy UAVs was a game-changer, allowing for rapid exploration of areas inaccessible to ground robots, such as shafts and high ledges.
- Tracked UGVs Excel in Rough Terrain: The paper highlights the surprising capability of rugged tracked platforms for challenging terrain, influencing the team's platform selection and design choices.
- Autonomy Over Communication: The team's experience consistently showed that prioritizing robotic autonomy and coordinated fleet behavior was more impactful than developing complex communication systems, especially in the face of unreliable networks.
Key Results
- Most Accurate Artefact Detection: The team won the "most accurate artefact detection" award at the Urban Event, with an error of 0.22 m.
- Robust Exploration: The system demonstrated successful autonomous exploration and mapping in diverse environments, including tunnels, urban structures, and caves.
- Decentralized SLAM Performance: The Wildcat SLAM solution achieved an error of < 0.05% over distance traveled during the Tunnel circuit.
- Effective Multi-Agent Coordination: The MRTA framework enabled agents to coordinate tasks and share information effectively, even when separated.
- Successful Deployment Across Events: The system was deployed and tested across multiple DARPA SubT events (STIX, Tunnel Circuit, Urban Circuit, and a custom Cave event), with continuous improvements driven by field results.
Limitations and Future Work
- Narrow Doorway Weakness: The approach, particularly with larger UGVs, struggled with very narrow doorways and small rooms, which were not well-represented in their training environments.
- Conservative Navigation: In some instances, autonomous navigation was too conservative, preventing traversal of obstacles that the robots were physically capable of overcoming.
- Operator Interface Complexity: While powerful, the operator interface for task prioritization (e.g., priority regions) was noted as being complex and requiring further simplification.
- Future Work: The authors plan to incorporate terrain risk into frontier selection and global navigation, improve coordination of exploration, introduce behaviors for passing through narrow openings and handling being stuck, enhance operator situational awareness with live colorized point clouds, improve UAV perception, develop smarter drop nodes, and generally enhance platform robustness. The focus on autonomy, coordination, and platform heterogeneity will continue for future deployments.
Summary updated