Abstract
This chapter introduces two kinds of motion path planning algorithms for mobile robots or unmanned ground vehicles (UGV). First, we present an approach of trajectory planning for UGV or mobile robot under the existence of moving obstacles by using improved artificial potential field method. Then, we propose an I-RRT* algorithm for motion planning, which combines the environment with obstacle constraints, vehicle constraints, and kinematic constraints. All the simulation results and the experiments show that two kinds of algorithm are effective for practical use.
Introduction
Mobile robot navigation faces two persistent challenges that limit the effectiveness of standard algorithms. The first is the Local Minimum Trap in traditional Artificial Potential Field (APF) methods. Standard APF treats the goal as an attractor and obstacles as repulsors; however, this often leads to "deadlocks" where opposing forces cancel out, or failures when an obstacle is positioned too close to the goal. The second challenge is Kinematic Ignorance in sampling-based algorithms like RRT*. These often produce mathematically optimal paths that are physically impossible for a real vehicle to follow, such as requiring zero-radius turns. Furthermore, these algorithms can converge too slowly in cluttered 3D environments.
This research proposes two major upgrades: an Improved APF that uses a genetic trust region to escape local minima, and an Improved RRT (I-RRT) that incorporates nonholonomic vehicle constraints for realistic pathfinding.
Methodology
1. Improved Artificial Potential Field (APF)
The authors transition from simple force vectors to potential field intensity. To solve the "Goal Unreachable" problem, they introduce a repulsive function with a coefficient that drops to zero as the robot approaches the goal.
The core innovation is the Genetic Trust Region. To escape local minima, the algorithm searches an annular (ring-shaped) region around the robot to identify the next "sub-goal." This search combines a genetic algorithm with a trust-region method, allowing the robot to "look ahead" and navigate out of traps.
As shown in the figure above, by restricting the search for the next step to a specific ring ($R \in [2V_{max}t_0/3, V_{max}t_0]$), the algorithm ensures the robot maintains a stable, efficient velocity without the computational burden of re-calculating the entire global map at every step.
2. Improved RRT (I-RRT)
The I-RRT algorithm integrates the robot's specific movement model (such as minimum turning radius) directly into the sampling process. It utilizes a Dual-Tree Sampling approach where two trees grow toward each other to speed up discovery. To maintain global optimality while improving convergence speed, I-RRT employs a "ball region" connection strategy, which acts as a partial greedy heuristic without falling into local optima.
Key Results
The proposed methods were evaluated in both Gazebo simulations and a physical laboratory setup using an omnidirectional mobile robot equipped with ultrasonic and infrared sensors.
- APF Reliability: In physical experiments, the improved APF reached targets in approximately 30 seconds with zero deadlock incidents, demonstrating a smooth field strength decrease near the goal.
- I-RRT* Performance: In 3D narrow-passage simulations, I-RRT significantly outperformed its predecessors. It achieved an average completion time of ~31.4 seconds, compared to ~115.9 seconds for B-RRT and ~218.5 seconds for standard RRT. Notably, I-RRT maintained a 0% failure rate, whereas standard RRT* failed nearly 20% of the time in narrow passages.
Conclusion
By building kinematics directly into the expansion step, I-RRT* produces "driveable" paths by default, eliminating the need for post-processing smoothing. While these methods show great promise, future work is needed to address the computational overhead of genetic trust regions on low-power hardware and to extend the models to multi-robot coordination and high-speed dynamics involving mass and inertia.
Summary updated