Robotics Lab URJC

Week 9 - Enhanced CARLA Performance and D.A.G.G.E.R. Dataset Implementation

November 25, 2025

Improved simulation stability and advanced dataset aggregation for robust autonomous driving

This week marks significant progress in both technical performance with the CARLA simulator and the implementation of the D.A.G.G.E.R. (Dataset Aggregation) methodology for training data generation.

We have successfully improved our technical workflow with CARLA, achieving stable driving sessions in Towns 01, 02, and 04 without crashes, while utilizing "Epic" quality settings for high-definition image capture. This enhancement significantly increases visual detail and consequently introduces greater variability into our training data. Using this improved setup, we generated a comprehensive dataset comprising 178,000 images representing key lane-following maneuvers: forward driving and lane recovery from the left lane back to the right lane.

However, during autonomous driving tests, we observed a critical limitation: the vehicle demonstrates uncertainty when approaching the right side of multi-lane roads, particularly evident in Town 04. This issue frequently manifests when encountering right-turn lanes and dividers. This observation raises an important question: Could the driving style (aggressive versus defensive) of the human expert provide better response velocities in these challenging scenarios?

Our second major task involved preparing a specialized dataset using the D.A.G.G.E.R. methodology. We implemented an approach that generates controlled steering perturbations through pulse bursts, modifying the vehicle's steering according to the following equation:

steering = raw_steering × (1 - δ(t-τ)) + 0.5 × δ(t-τ)

In this implementation, the delta function δ(t-τ) corresponds to pulses of variable duration between 0.5 and 1.0 seconds. The term 0.5 × δ(t-τ) initially limits steering modifications to two discrete values: +0.5 and -0.5.

The construction of this dataset has proven particularly challenging due to the inherent randomness that can force the vehicle into frontal collisions with obstacles. This necessitates the development of a more systematic methodology for perturbation application to ensure meaningful training examples while maintaining vehicle safety during data collection.