Ian Utnehmer

I'm a Software Engineer specializing in computer vision / ML for Robotics in my free time. My current project has finished its first iteration.

Autonomous Scout is my thesis project: a low-cost robotics navigation system that closes the gap between 2D LiDAR-based obstacle avoidance and real-world terrain awareness. I built the system around a Jetson Nano/JetRacer platform using Python, PyTorch, ROS, and TensorRT, combining LiDAR SLAM, camera-based perception, and autonomous path planning.

The project detects hazards that standard 2D LiDAR often misses, including ground-level paper and terrain such as grass. I developed a camera segmentation stack for constrained edge compute, including swap-space mitigation, tooling for CSI camera hangs, and a labeled-image pipeline for training and deployment. I first deployed a ResNet-18 model for paper detection, then reused the same segmentation pipeline with a Cityscapes SegNet-18 model to extend the system toward terrain detection.

I integrated monocular depth estimation with Monodepth2 to approximate distance from RGB imagery alone, then projected detected paper and terrain regions into a 2D ROS costmap. This allowed the robot to reason about hazards LiDAR could not directly see and convert visual detections into navigation constraints.

I fused camera detections with LiDAR-based navigation by separating live obstacle handling from persistent mapping logic, exposing results in RViz and feeding them into move_base for autonomous path planning. The full stack successfully avoided paper in initial tests, then avoided grass after plugging the Cityscapes SegNet-18 model into the same pipeline.

I also built the supporting infrastructure for reliable deployment, including rosbridge-based host control, teleoperation and move-to-point scripts, static-IP networking through a NetworkManager dispatcher script, and localization debugging around AMCL drift and base_footprint-to-laser_frame transforms. Overall, the project demonstrates a practical perception-and-planning system that improves navigation safety, consistency, and readiness for indoor and outdoor terrain navigation.