TL;DR
This paper presents a fast, real-time system for autonomous navigation in unknown off-road environments. It combines lidar-based mapping with an A* planner and a novel kinematic trajectory sampling approach to generate safe and feasible paths at high speeds, outperforming traditional methods in challenging terrain.
What Problem Does This Paper Address?
Navigating off-road environments autonomously is significantly harder than on-road navigation. The terrain is unstructured, unpredictable, and can present unique challenges. For instance, what appears to be an impassable swamp might be drivable, while tall grass or small bushes might be perceived as obstacles but can be driven over. Traditional methods often struggle with these nuances, either being too slow for real-time operation, unable to handle non-holonomic constraints of vehicles, or failing to adapt to dynamic changes in the environment. This paper tackles the problem of creating a robust, fast, and adaptable local mapping and planning system that can operate at the vehicle's performance limits in such unknown terrains.
Methods
The proposed system integrates several key components:
- Mapping: A lidar (VLP-32c) is used to build a local map of the environment. This map is a composite of obstacle detection and terrain gradient information. An Extended Kalman Filter (EKF) fuses odometry data with GPS/IMU for accurate vehicle localization. The mapping module operates at 10 Hz.
- Planning:
- A* Planner: This algorithm is used for medium-range planning, finding an optimal path across the generated cost map. It's chosen for its speed and optimality guarantee within the map. The A* planner runs at 5 Hz.
- Kinematic Trajectory Sampling: To ensure kinematically feasible paths, the system samples a set of control inputs (angular velocity, durations for turns and straight segments) over a short time horizon. This generates approximately 1000 kinematically feasible trajectories.
- Path Selection: Each generated trajectory is evaluated using a cost function that considers the map cost, distance to the goal, heading error, and path curvature. The trajectory with the lowest cost is selected and sent to the vehicle controller. This trajectory sampling and selection module operates at a high rate of 30 Hz.
The novelty lies in the combination of these elements, particularly the kinematic trajectory sampling that explicitly generates feasible paths and the cost map that incorporates both obstacle and gradient information for better off-road performance.
Datasets
The paper does not explicitly mention or introduce new datasets. Instead, it describes experiments conducted using a Clearpath Robotics Warthog vehicle equipped with a VLP-32c lidar and a VectorNav VN-300 IMU/GPS. The tests were performed at the Texas A&M Rellis Campus, utilizing natural off-road terrain with some added obstacles like traffic cones.
Key Insights and Ideas
- Composite Cost Map is Crucial for Off-Road: Simply detecting obstacles isn't enough for off-road navigation. Incorporating terrain gradients allows the vehicle to intelligently decide whether to drive over smaller obstacles or navigate around them, leading to more efficient and adaptable paths.
- Trajectory Sampling for Feasibility: While A* provides an optimal path on the map, it doesn't guarantee the vehicle can actually follow it due to kinematic constraints. Sampling kinematically feasible trajectories and selecting the best one ensures the robot can execute the planned path safely.
- High-Frequency Trajectory Planning for Responsiveness: Operating the trajectory sampling and selection at 30 Hz allows the system to react quickly to new sensor information and adapt its path in real-time, crucial for dynamic off-road environments.
- Balancing Optimality and Feasibility: The system effectively balances the optimality of the A* path with the kinematic feasibility ensured by the trajectory sampler, leading to robust performance.
Key Results
The system demonstrated real-time operation at speeds of 3 m/s, close to the vehicle's maximum speed of 4.5 m/s. Experiments showed the system successfully navigating:
- A relatively flat field with traffic cones, where the vehicle adapted its path when it overshot the initial A* plan.
- An approximately 50 m off-road trail with tall grass and brush, where minor deviations were made based on updated terrain gradient information.
The paper highlights that their trajectory sampling approach avoids collisions that a pure pursuit controller might encounter in similar scenarios.
Limitations and Future Work
- Greedy Global Planning: The current system focuses on local planning. If the target point is outside the Region of Interest (ROI) of the local map, the global path produced is described as "greedy," implying it might not be optimal or safe in the long term.
- Integration with Global Planner: A significant future direction is integrating this local planner with a global planner to provide more comprehensive navigation.
- Semantic Segmentation: The authors plan to incorporate semantic segmentation to identify moving obstacles and improve planning around them.
- Moving Obstacle Handling: The current system is designed for static environments. Handling dynamic obstacles is a key area for future development.
Summary updated