Introduction
Autonomous exploration presents a fundamental "chicken and egg" challenge: a robot requires a map to plan its trajectory, yet it must move through the environment to construct that very map. In complex 3D spaces, robots must simultaneously update their world representation and identify "frontiers"—the boundaries between known and unknown space. Existing approaches often face a trade-off between high-resolution local planning, which is computationally intensive and myopic, and global planning, which lacks the precision for safe, high-speed maneuvering.
TARE (Technologies for Autonomous Robot Exploration) addresses this by introducing a hierarchical planning framework. This system maintains a dense map for immediate local maneuvering and a sparse map for global connectivity, achieving an 80% increase in exploration efficiency while significantly reducing computational overhead.
Hierarchical Planning Methodology
The core of TARE is its division of the exploration task into two distinct layers that operate at different scales and resolutions.
Local Planning (The "Micro" View)
Within a defined "local planning horizon"—a bounding box surrounding the robot—the system maintains a high-resolution representation. It employs iterative random sampling to identify a minimal set of viewpoints covering all nearby "uncovered surfaces." These viewpoints are sequenced by solving a Traveling Salesman Problem (TSP) and smoothed into a kinodynamically feasible path using cubic splines. This ensures the robot explores its immediate vicinity with high detail and safety.
Global Planning (The "Macro" View)
Outside the local horizon, the environment is partitioned into large cuboid subspaces. TARE maintains a sparse roadmap connecting the centroids of these subspaces. A global path is computed to link the robot’s current position to distant, unexplored regions. By connecting local and global paths at the boundary of the planning horizon, the robot avoids getting trapped in local minima and ensures systematic coverage of the entire environment.
Experimental Evaluation
The authors validated TARE through extensive simulation and real-world hardware tests:
- Simulation: Large-scale outdoor university campus environments for aerial robotics.
- Real-World Ground Robotics: A complex four-story parking garage and adjacent patio.
- DARPA Subterranean Challenge: Field testing at the Satsop Nuclear Plant as part of the CMU-OSU team's entry.
To support further research, the authors released a benchmark software environment containing simulation models and navigation modules at cmu-exploration.com.
Key Results and Performance
TARE demonstrates significant improvements over state-of-the-art methods such as NBVP (Next-Best-View Planner) and GBP (Graph-Based Planner).
- Exploration Efficiency: TARE achieved an 80% increase in volume explored per second compared to competitors.
- Computational Cost: The framework consumed less than 50% of the computation time of existing methods. In aerial tests, TARE’s runtime averaged ~0.21s per re-plan, whereas other planners required between 4.6s and 7.4s.
- Robustness: In real-world multi-story environments, TARE was the only method to achieve full exploration; other planners either timed out or failed to navigate complex structures like spiral driveways.
Conclusion
TARE proves that hierarchical decomposition is a powerful tool for autonomous exploration. By focusing computational resources on the immediate vicinity while maintaining a coarse global strategy, robots can explore complex 3D environments with unprecedented speed and efficiency. Future work includes integrating platform-specific motion models and adapting the framework for dynamic environments with moving obstacles.
Summary updated