Autonomous Obstacle Avoidance on the GEM e4

CS588: Autonomous Vehicle System Engineering

A Highbay navigation framework for the Polaris GEM e4 that performs obstacle-aware planning and control. An RGB-D + YOLO perception module estimates obstacle positions, a hierarchical Frenet-frame replanner samples safe detours, and a low-level Stanley controller tracks the chosen path. The system is validated first in high-fidelity simulation and then on the physical GEM e4.

Simulation

In simulation, obstacles (including a pedestrian added to the scene) are detected from the front-camera stream and avoided in closed loop, with a mean tracking error of 0.41 m (RMSE 0.66 m).

Hardware Deployment

The same planning and tracking pipeline transfers to the physical GEM e4: the vehicle follows the reference, executes a lateral detour around the obstacle, and returns to path with a minimum clearance of 1.84 m (lateral RMSE 0.38 m).