The thesis involves implementing the MILE model (Model-based Imitation LEarning for urban driving), trained on the CARLA simulator, into the Duckietown environment to evaluate its performance in navigating unprotected intersections.

Experiments were conducted using the Gym-Duckietown simulator, where Duckiebots navigated a 4-way intersection across multiple trajectories. Metrics such as success rate, drivable area compliance, and ride comfort were used to assess performance.

The findings indicate that while the MILE model achieved state-of-the-art performance in the CARLA simulator, its generalization to the Duckietown environment without additional training was, as probably expected due to the sim2real gap, limited.

The BEVs generated by MILE were not sufficiently representative of the actual road surface in Duckietown, leading to suboptimal navigation performance. In contrast, the homographic BEV method, despite its assumption of a flat world plane, provided more accurate representations for intersection navigation in this context.

As for most approaches in robotics, there are limitation and tradeoffs to analyze.

Here are some technical challenges of the proposed approach:

• Generalization across environments: one of the challenges is ensuring that the 3D image feature representation generalizes well across different simulation environments, such as Duckietown and CARLA. The differences in scale, road structures, and dynamics between simulators can impact the performance of the navigation system.
• Accuracy of BEV representations: the transformation of camera images into Bird’s Eye View (BEV) representations has reduced accuracy, especially when dealing with low-resolution or distorted input data.
• Real-time processing: the integration of 3D image features for navigation requires substantial computational resources with respect to utilizing 2D features instead. Achieving near real-time processing speeds for navigation tasks such as intersection navigation, is challenging.