Start Date
1-5-2026 12:00 PM
End Date
1-5-2026 1:00 PM
Description
Autonomous delivery robots require robust, real-time navigation systems to safely traverse dynamic outdoor environments like college campuses. This project presents an edge-optimized vision pipeline deployed on a resource-constrained Jetson Nano-based robot to accurately identify drivable surfaces and detect path intersections. We employ a lightweight semantic segmentation Convolutional Neural Network (LRASPP MobileNetV3-Large), initially pretrained on the Cityscapes dataset. By applying transfer learning to a manually refined, locally labeled dataset of our campus, we significantly improved feature extraction for urban textures and adapted the model to our specific use case. To ensure efficient edge deployment, the network is optimized using TensorRT [1], achieving inference speeds of approximately 15 FPS.
To translate these segmentations into navigational decisions, we developed a novel Voronoi-based split detection algorithm. This technique transforms a 2D binary mask of the drivable surface into a 1D skeletonized representation [2], effectively distilling the shape into its essential topological structure. The resulting skeleton serves as the primary input for classification logic to determine whether the robot is currently encountering a ”split” in its path. While traditional methods for identifying intersections often rely on global self-attention [3], scene-level CNN labels [4, 5], or top-down occupancy grids [6], our solution implements a non-traditional application of Voronoi Diagrams specifically designed to identify path splits for local navigation.
Experimental results demonstrate that this Voronoi-based approach outperforms traditional methods (such as IPCA and Histogram classifiers) on the noisy Cityscapes dataset and achieves 95% accuracy on our custom campus dataset. This research provides a computationally efficient and highly accurate alternative for local navigation, contributing to the broader goal of developing low-cost autonomous delivery systems.
Last-Mile Autonomous Delivery Robot
Autonomous delivery robots require robust, real-time navigation systems to safely traverse dynamic outdoor environments like college campuses. This project presents an edge-optimized vision pipeline deployed on a resource-constrained Jetson Nano-based robot to accurately identify drivable surfaces and detect path intersections. We employ a lightweight semantic segmentation Convolutional Neural Network (LRASPP MobileNetV3-Large), initially pretrained on the Cityscapes dataset. By applying transfer learning to a manually refined, locally labeled dataset of our campus, we significantly improved feature extraction for urban textures and adapted the model to our specific use case. To ensure efficient edge deployment, the network is optimized using TensorRT [1], achieving inference speeds of approximately 15 FPS.
To translate these segmentations into navigational decisions, we developed a novel Voronoi-based split detection algorithm. This technique transforms a 2D binary mask of the drivable surface into a 1D skeletonized representation [2], effectively distilling the shape into its essential topological structure. The resulting skeleton serves as the primary input for classification logic to determine whether the robot is currently encountering a ”split” in its path. While traditional methods for identifying intersections often rely on global self-attention [3], scene-level CNN labels [4, 5], or top-down occupancy grids [6], our solution implements a non-traditional application of Voronoi Diagrams specifically designed to identify path splits for local navigation.
Experimental results demonstrate that this Voronoi-based approach outperforms traditional methods (such as IPCA and Histogram classifiers) on the noisy Cityscapes dataset and achieves 95% accuracy on our custom campus dataset. This research provides a computationally efficient and highly accurate alternative for local navigation, contributing to the broader goal of developing low-cost autonomous delivery systems.