Introduction
Developing and deploying autonomous robots often involves significant engineering overhead to integrate disparate modules like state estimation, terrain mapping, and collision avoidance. Researchers at Carnegie Mellon University have addressed this challenge by introducing an open-source software stack designed to bridge the gap between high-level planning research and real-world deployment. This "Autonomous Exploration Development Environment" provides a standardized middleware that abstracts hardware complexities while offering high-fidelity simulation and robust local planning.
System Architecture
The system is built upon a three-tier architecture designed for seamless sim-to-real parity. By ensuring that the same code runs in both simulation and on physical hardware, the environment eliminates the common pitfalls of algorithms that rely on ground-truth data unavailable in the real world.
- Bottom Layer: Manages low-level state estimation and motion control.
- Mid-Layer (The Development Environment): The core contribution, featuring a local planner that utilizes Monte Carlo sampling for collision-free paths. It includes a terrain analysis module that builds voxel-based cost maps and identifies "negative obstacles" like pits or cliffs by analyzing data point distributions.
- Top Layer (High-Level Planners):
- TARE (Exploration): A hierarchical planner that solves the Traveling Salesman Problem at both global and local levels to optimize coverage efficiency.
- FAR (Route Planning): A visibility graph-based planner that dynamically updates maps and recalculates routes in less than 0.3 milliseconds.
Simulation Environments and Datasets
To stress-test navigation across diverse challenges, the paper provides five distinct simulation environments: Campus (undulating terrain), Indoor (narrow corridors), Garage (multi-floor with steep ramps), Tunnel (complex underground networks), and Forest (cluttered outdoor space).
The environment is also compatible with photorealistic models from Matterport3D and AI Habitat. Real-world validation was conducted using data from the DARPA Subterranean Challenge, where the system demonstrated its ability to handle extreme, unstructured environments.
Key Results and Performance
The effectiveness of the stack was proven during the DARPA Subterranean Challenge, where the CMU-OSU team’s robot—powered by this system—explored 26 out of 28 sectors, the highest completion rate in the competition.
In terms of efficiency, the TARE planner successfully explored a four-story garage (1,839 meters) in approximately 31 minutes. The FAR planner showed remarkable computational efficiency, requiring only 15% of a single CPU thread for its visibility graph expansions. These results highlight the system's ability to provide high-level strategic planning without sacrificing real-time responsiveness.
Limitations and Future Work
Currently, the system is optimized for differential-drive or skid-steer kinematics and does not natively support car-like (Ackermann) steering constraints. Additionally, while the simulation supports vision-based research, the core navigation modules are primarily built for 3D Lidar data. Future developments aim to include multi-robot coordination and "mixed-initiative" navigation, allowing robots to execute high-level natural language instructions from human operators.
Summary updated