Source-linked AI summary
Autonomous Exploration Development Environment and the Planning Algorithms
Chao Cao, Hongbiao Zhu, Fan Yang, Yukun Xia, Howie Choset, Jean Oh, Ji Zhang
TL;DR
The paper addresses the need for resources to develop high-level planning algorithms and integrate autonomous navigation systems. It presents an open-source development environment together with TARE and FAR planners, then reports their use in the CMU-OSU Team’s DARPA Subterranean Challenge navigation system, which explored 26 of 28 sectors.
Problem
The paper aims to provide resources for developing high-level planning algorithms and integrating complete autonomous navigation systems.
Source: §I · p. 1, §III · p. 2Method
The authors combine an environment with models and navigation modules, TARE for exploration, and FAR for route planning into a geometry-based navigation stack.
Source: §I · p. 1, §III · p. 2, A. TARE Planner for Exploration · p. 3Results
The CMU-OSU Team explored 26 of 28 sectors and received the “Most Sectors Explored Award” using the repositories as its main navigation system.
Source: §I · p. 1Takeaways & Limitations
The repositories serve as a platform for developing and deploying ground-vehicle navigation systems, with potential applications in simulated and real systems.
Source: §I · p. 1The system is scoped to ground vehicles rather than aerial vehicles; the authors identify multi-robot coordination and exploration as future extensions.
Source: §VIII · p. 6
Abstract
from arXiv · showhide
Autonomous Exploration Development Environment is an open-source repository released to facilitate the development of high-level planning algorithms and integration of complete autonomous navigation systems. The repository contains representative simulation environment models, fundamental navigation modules, e.g., local planner, terrain traversability analysis, waypoint following, and visualization tools. Together with two of our high-level planner releases -- TARE planner for exploration and FAR planner for route planning, we detail usage of the three open-source repositories and share experiences in the integration of autonomous navigation systems. We use DARPA Subterranean Challenge as a use case where the repositories together form the main navigation system of the CMU-OSU Team. In the end, we discuss a few potential use cases in extended applications.
I. INTRODUCTION
The paper presents an open-source development environment with TARE and FAR planners as a geometry-based navigation stack for ground vehicles. The repositories support developing high-level planners, integrating them into a three-tier system, and moving them from simulation toward vehicle deployment.
I. INTRODUCTION
- The environment, TARE planner, and FAR planner form a geometry-based navigation stack for ground vehicles.The repositories are intended as a platform for developing and deploying advanced navigation systems.Source: §I · p. 1
- The development environment supplies environment models and fundamental modules, including local planning, terrain traversability analysis, and waypoint following.It serves as the mid-layer between state estimation and motion control below and high-level planning above.Source: §I · p. 1, Figure 1 · p. 1
- The system avoids simulation-only terrain-segmentation ground truth by computing traversability from range data, as in a real system.The intended workflow is to develop planners and migrate the repositories to a vehicle computer.Source: §I · p. 1
II. RELATED WORK
The related work spans datasets, simulation environments, and integrated navigation systems. The paper positions its contribution as an open-source development environment that complements these resources with planning support and interfaces for system integration.
II. RELATED WORK
- Datasets provide real sensor readings for developing and benchmarking perception and planning, particularly for users without integrated sensor suites.The section names outdoor and indoor datasets, including KITTI, NYU-Depth, and TUM SLAM.Source: §II · pp. 1–2
- Simulation environments provide ground truth, such as vehicle pose and semantic segmentation, and support closed-loop tests of complete navigation systems.Examples include Carla and AirSim for driving and flying, and indoor environments such as AI Habitat, [25].Source: §II · p. 2
- Existing system-level efforts include ROS Navigation Stack and open-source aerial-navigation packages from ETH and HKUST.The cited aerial packages cover components from sensor drivers and state estimation to planning and exploration.Source: §II · p. 2
- The paper’s development environment provides interfaces between state estimation, motion control, and high-level planning, with modules and models for ground-vehicle navigation.It also supports Matterport3D house models through an AI Habitat interface, [25].Source: §II · p. 2
III. DEVELOPMENT ENVIRONMENT
The development environment is a platform for developing and benchmarking high-level ground-vehicle planners, with five environment models, fundamental navigation modules, and visualization and debugging tools.
III. DEVELOPMENT ENVIRONMENT
- The environment provides five models alongside fundamental navigation modules and visualization and debugging tools.Its stated purpose is developing and benchmarking high-level planning algorithms for ground-vehicle navigation.Source: §III · p. 2
A. Environment Models
- The five models cover a campus, indoor corridors, a five-floor garage, a tunnel network, and a cluttered forest setting.Their settings include undulating terrain, narrow corridors, sloping multi-floor terrain, connected tunnels, and trees with houses.Source: A. Environment Models · p. 2
- Matterport3D house models are supported with scan data and AI Habitat-rendered RGB, depth, and semantic images.Fig. 3 depicts a house model and its corresponding rendered images.Source: A. Environment Models · p. 2, Figure 3 · p. 3
B. Local Planner
The local planner guides the vehicle to high-level waypoints using collision-free paths, selecting motion primitives according to obstacles and waypoint direction.
B. Local Planner
- The planner precomputes motion primitives associated with nearby 3D locations and detects obstacle collisions in milliseconds.The primitives are modeled as Monte Carlo samples and organized in groups.Source: B. Local Planner · p. 3
- It selects the group of motion primitives with the maximum likelihood toward the waypoint after identifying collisions.The selected primitives form collision-free paths leading to the waypoint.Source: B. Local Planner · p. 3
- Terrain analysis supplies the terrain map used to determine ground-vehicle traversability, and the planner can optionally accept additional range data for collision avoidance.Source: B. Local Planner · p. 3
C. Terrain Traversability Analysis
Terrain traversability analysis estimates local ground shape and converts it into traversal costs and traversability labels. Its maps also account for negative obstacles when that option is enabled.
C. Terrain Traversability Analysis
- A voxel grid and adjacent-voxel point distributions estimate ground height, while local terrain smoothness determines traversal cost.Points farther from the ground receive higher traversal costs.Source: C. Terrain Traversability Analysis · p. 3
- Fig. 4 places the vehicle and coordinate frame at the same location and shows collision-free motion primitives as red paths.Source: Figure 4 · p. 3
- Fig. 5 maps a 40m x 40m area, marking traversable points green and non-traversable points red; with negative-obstacle handling on, empty-data areas are non-traversable.Source: C. Terrain Traversability Analysis · p. 3, Figure 5 · p. 3
D. Visualization and Debugging Tools
The section describes visualization and debugging tools alongside high-level planning modules. TARE uses hierarchical, multi-resolution planning, and its evaluations report improved exploration and computation efficiency over comparison methods.
D. Visualization and Debugging Tools
- Visualization tools display maps, explored areas, and trajectories, and log explored volume, travel distance, and runtime; joystick control and operation modes support debugging.Source: D. Visualization and Debugging Tools · p. 3
IV. HIGH-LEVEL PLANNERS
- High-level planners generate task-specific waypoints at lower frequencies, while local navigation modules follow routes and react to obstacles.Source: §IV · p. 3
A. TARE Planner for Exploration
- TARE plans globally with low-resolution information and locally with high-resolution information, optimizing the overall exploration path by solving a traveling salesman problem at each level.The local detailed path and global coarse paths are connected to form the exploration path.Source: A. TARE Planner for Exploration · p. 3, Figure 6 · p. 4
- TARE demonstrates significantly higher exploration and computation efficiency than NBVP, GBP, and MBP in evaluations across large, complex simulated and real environments.A representative four-storage-garage and patio exploration covered the environment after 1839m of travel in 1907s.Source: A. TARE Planner for Exploration · pp. 3–4
B. FAR Planner for Route Planning
FAR dynamically builds a reduced visibility graph for route planning in known and unknown environments. The supplied evaluation reports fast path finding and replanning over long distances.
B. FAR Planner for Route Planning
- FAR models obstacles as polygons, merges them across sensor frames, and uses their edge points to develop a reduced visibility graph dynamically during navigation.Source: B. FAR Planner for Route Planning · p. 4
- A path is found within 0.3ms in all experiments, and FAR demonstrates fast replanning over long distances.The planner is compared with RRT*, RRT-connect, A*, and D* Lite.Source: B. FAR Planner for Route Planning · p. 4
- In unknown environments, FAR registers obstacles and replans when it discovers a better route; in known environments, it plans from a prior map’s visibility graph.The illustrated routes use a Matterport3D house model.Source: B. FAR Planner for Route Planning · p. 4, Figure 8 · p. 4
V. BEST PRACTICES
The best practices describe how waypoint distance and look-ahead selection affect local planning, and how dynamic-obstacle removal is bounded by sensor data conditions.
V. BEST PRACTICES
- Keep a waypoint at least 3.75m away by default when the vehicle is not expected to stop there; reducing the distance can help find collision-free primitives through narrow openings.The local planner’s collision checking does not consider obstacles beyond the waypoint.Source: §V · p. 4
- For turns of at least 90 degrees, a poorly set look-ahead distance can make the waypoint jump behind the vehicle and cause oscillation; choose a visible waypoint on the path’s starting segment.Source: §V · pp. 4–5
- Dynamic-obstacle removal is limited to within 5m because range data becomes sparse farther away and balancing obstacle removal against preserving thin structures is difficult.Task-specific dynamic-obstacle handling is advised; both TARE and FAR include a removal step.Source: §V · p. 5
VI. SYSTEM INTEGRATION
The development environment provides a mid-layer between state estimation and motion control, with interfaces for high-level planners. Its integration with TARE and FAR supports CMU-OSU Team exploration and route planning in the DARPA Subterranean Challenge.
A. Basic System
- The basic system places the development environment between state estimation and motion control, connecting high-level planners through terrain maps, waypoints, navigation boundaries, and speed.This setup lets users develop high-level planners without needing to understand interfaces among lower-level navigation modules.Source: A. Basic System · p. 5, Figure 10 · p. 5
- The system supports differential-drive platforms, while omnidirectional and legged robots do not exploit lateral movement and car-like platforms are unsupported.It accepts registered world-frame scans and can also provide sensor-frame scans with synchronized state estimation.Source: A. Basic System · p. 5
B. DARPA Subterranean Challenge System
- The competition system extends the basic system with loop-closure adjustments for global planners and local modules operating in the odometry frame.Local planning and terrain analysis were extended for negative obstacles, while TARE and FAR ran alongside recovery and coordination modules.Source: B. DARPA Subterranean Challenge System · pp. 5–6, Figure 11 · p. 5
- In a representative Urban Circuit run, TARE explored the floor autonomously over 886m in 1458s.The run took place at Satsop Nuclear Plant, Washington.Source: Figure 12 · p. 6
- Three vehicles used TARE and FAR across the final competition, traveling 596.6m, 499.8m, and 445.2m over 2259s.TARE handled exploration and FAR handled requested relocations between areas.Source: B. DARPA Subterranean Challenge System · p. 6
VII. EXTENDED APPLICATION EXAMPLES
The system is presented as a platform for multidisciplinary navigation research, with examples spanning visual learning, human instructions, and social navigation. Its simulation and semantic resources support these applications.
Extended application examples
- The final-competition example pairs FAR transit segments with TARE exploration segments across vehicle trajectories.The red and blue paths include FAR relocation, while other trajectory segments use TARE to explore.Source: Figure 13 · p. 6
Self-supervised learning for visual navigation
- The system can collect training data for vision-based navigation using Lidar-based self-supervision, with Matterport3D models supporting simulation-to-real adaptation.The described test-time method navigates using only a camera.Source: §VII · p. 6
Mixed-initiative navigation with human instructions
- Matterport3D semantic ground truth can help locate a door for natural-language navigation, while reference trajectories train a method that does not access that ground truth.The example instruction is to go through the front door and then turn left.Source: §VII · p. 6
Social navigation in crowded environments
- Gazebo allows custom human avatars and provides pose ground truth in simulation for developing social navigation algorithms.The ground truth can potentially simplify development by bypassing human pose detection in simulation.Source: §VII · p. 6
VIII. DISCUSSION AND FUTURE ACTIVITIES
The authors aim to provide a generic navigation system for multidisciplinary research, choosing ground vehicles for their sensor-computer payload capacity and battery life. They plan to extend the system to multi-robot coordination and exploration and organize tutorials.
Discussion and future activities
- Ground vehicles were chosen over aerial vehicles for their capacity to carry heavy sensor-computer payloads and extended battery life.The authors connect this choice to contemporary AI research involving sophisticated sensors, computers, and powerful GPUs.Source: §VIII · p. 6
- Planned future activities include extending the system to multi-robot coordination and exploration and organizing tutorials at ICRA and CVPR.The tutorials are intended to help users apply the system in their work.Source: §VIII · p. 6