3.4.1 Comparative Evaluation of Path Planning Strategies
Real-time autonomous navigation experiments were carried out in the four scenarios in Fig. 3, where the Improved LeGO-LOAM facilitated localization and mapping. Our primary focus was on evaluating the impact of the local planner on the resultant trajectory. The relevant parameters are also displayed in Table 2. To visually compare the effects of algorithmic improvements, the comparative results are presented on a two-dimensional grid map. As shown in Fig. 6, the original OpenPlanner is represented by the blue line, whereas our Enhanced OpenPlanner is indicated by the red line. These trajectories are derived based on the odometer information of the robot.
Fig. 6 (a) and (e) display the results of navigation and obstacle avoidance experiments in the crop row scenario. The lower left of Fig. 6 (a) provides the location and details of the obstacle. The plant edges comprise leaves close to the ground, posing a challenge for successful robot avoidance. The comparison results in Fig. 6 (e) show that both algorithms avoid the obstacles, and their trajectories initially appear almost identical. However, a difference emerges when encountering obstacles: our algorithm executes a larger-angle evasion due to the consideration of plant density, whereas the robot relying on the original OpenPlanner algorithm encounters the edge of the plants. Therefore, in this scenario, our algorithm proves safer.
Fig. 6 (b) and (f) illustrate navigation and obstacle avoidance results within a landscape greenhouse. We configured a TurtleBot2 to shift laterally on the road as a dynamic obstacle. Fig. 6 (b) provides details of our algorithm’s avoidance of the dynamic obstacle, while Fig. 6 (f) contrasts both algorithms. The results reveal that both methods are capable of effectively avoiding dynamic obstacles and reaching the endpoint. However, a close examination reveals that the original OpenPlanner algorithm plans a more inward route to take a shorter path. Given the presence of branches stretching onto the road, this may cause the robot to collide with some branches. In contrast, the enhanced OpenPlanner algorithm prioritizes plant density over distance expenditure. This experiment demonstrates that the enhanced OpenPlanner algorithm is not only capable of avoiding dynamic obstacles but also ensures greater safety.
Fig. 6 (c) and (g) present navigation results within a challenging unstructured greenhouse. This is the most challenging scenario of this study, where both the rugged road surface and roadside vegetation pose significant challenges to navigation. As shown in the figures, the original OpenPlanner algorithm fails in this environment. Fig. 6 (c) presents the failed scenario and its cause. In local planning, the original OpenPlanner algorithm prioritizes distance cost, leading to traversal over many complex terrains. Eventually, the algorithm fails due to severe drift in pose estimation. In contrast, our algorithm, considering ground structure, safely reaches the target point. Even though it takes a longer route, this approach is meaningful from a practical perspective. This scenario underscores the crucial role of ground factors in successful navigation in complex environments.
Fig. 6 (d) and (h) aim to test navigation performance in low-lying plant environments. We set the target point at the edge of the low-lying plants. The LiDAR trajectory in Fig. 6 (d) reveals that the original OpenPlanner algorithm travels very close to the plants. This implies potential harm to low-lying plants if the robot is too large, as the algorithm does not consider plant density during local path planning. On the other hand, as seen in comparison Fig. 6 (h), our algorithm effectively keeps a safe distance from the low-lying plants, planning a more feasible and safer trajectory.
From the tests conducted across the four scenarios, it is evident that our enhanced OpenPlanner can effectively evade both static and dynamic obstacles during local path planning. Compared to the original OpenPlanner, the improved algorithm maintains a safer distance from the plants, which aids in protecting them from possible vehicle damage. This includes avoidance of plant branches protruding from the road edges and low-lying plants. Moreover, our algorithm proficiently avoids rugged terrain, contributing to a higher success rate in complex environment navigation. This robustness makes our algorithm more practical for real-world applications.