Terrain-aware path planning via semantic segmentation and uncertainty rejection filter with adversarial noise for mobile robots

Abstract

In ground mobile robots, effective path planning relies on their ability to assess the types and conditions of the surrounding terrains. Neural network-based methods, which primarily use visual images for terrain classification, are commonly employed for this purpose. However, the reliability of these models can vary due to inherent discrepancies between the training images and the actual environment, leading to erroneous classifications and operational failures. Retraining models with additional images from the actual operating environment may enhance performance, but obtaining these images is often impractical or impossible. Moreover, retraining requires substantial offline processing, which cannot be performed online by the robot within an embedded processor. To address this issue, this paper proposes a neural network-based terrain classification model, trained using an existing data set, with a novel uncertainty rejection filter (URF) for terrain-aware path planning of mobile robots operating in unknown environments. A robot, equipped with a pretrained model, initially collects a small number of images (10 in this work) from its current environment to set the target uncertainty ratio of the URF. The URF then dynamically adjusts its sensitivity parameters to identify uncertain regions and assign associated traversal costs. This process occurs entirely online, without the need for offline procedures. The presented method was evaluated through simulations and physical experiments, comparing the point-to-point trajectories of a mobile robot equipped with (1) the neural network-based terrain classification model combined with the presented adaptive URF, (2) the classification model without the URF, and (3) the classification model combined with a nonadaptive version of the URF. Path planning performance measured the Hausdorff distances between the desired and actual trajectories and revealed that the adaptive URF significantly improved performance in both simulations and physical experiments (conducted 10 times for each setting). Statistical analysis via t-tests confirmed the significance of these results.