Autonomous Driving Model Training Report

1. Enhanced Dataset Construction Methodology

Building upon the analysis performed in Week 25, this week's work focused on implementing controlled dataset generation techniques to increase the representation of extreme steering values and improve model generalization. The modifications applied to the original data collection script (based on Jorge Rodríguez's framework) introduced two critical enhancements:

• Controlled Dagger Triggering: Instead of random activation, the Dagger mechanism (simulating a drunk driver) is now manually triggered via a steering wheel button. This allows precise control over recording sessions, deliberately increasing the number of samples with extreme steering angles.
• Offline Tangent Transformation: Steering command values are processed using a tangent function to amplify the effect of steering inputs, saturating extreme values. This preprocessing step is intended to facilitate model training by accentuating the differences between steering commands.

Following this methodology, four distinct datasets were constructed to evaluate the impact of town diversity and class balancing:

Figure 1: Distribution of steering values by category obtained for Town01 only (13,500 images). The baseline dataset shows a concentration around straight driving values.

Figure 2: Balanced (weighted) distribution for Town01 only (15,000 images). Weighting provides greater representation across all steering categories.

Figure 3: Raw distribution for combined dataset (Towns 01, 02, 04, 12) with 37,000 samples. Greater town diversity increases overall variability.

Figure 4: Balanced (weighted) distribution for combined dataset (Towns 01, 02, 04, 12) with 41,000 images. Uniform representation across steering categories is achieved.

2. Experimental Results & Performance Evaluation

The models trained on the four datasets were evaluated on previously unseen environments to assess generalization capabilities. The most significant outcomes are summarized below:

Key Achievements

• Full circuit completion in Town07 (Generalization): The model trained on the balanced multi-town dataset (Figure 4) successfully completed the entire rural circuit in Town07, a scenario not included during training. This represents a major milestone in generalization capability.
• Successful 90° turns in urban environments: The agent demonstrated reliable execution of both left and right 90-degree turns in Town01, navigating urban intersections without exiting the drivable path.
• Enhanced stability: Oscillations observed in previous models were significantly reduced, resulting in smoother trajectory following.

3. Comparative Analysis: Dataset Strategies

The systematic comparison across the four datasets yielded critical insights into the effectiveness of balancing and multi-town training:

Key findings: The combination of multi-town data diversity and category-based balancing (weighting extreme steering values) proved essential for achieving generalization to unseen environments. The balanced multi-town dataset delivered superior results despite using only one-quarter of the sample count previously required, demonstrating that data quality and distribution are more critical than raw quantity.

4. Technical Analysis: Tangent Preprocessing & Balancing Effects

Two methodological innovations were central to this week's improvements:

Controlled Dagger Triggering

The ability to manually trigger Dagger events via a steering wheel button enabled targeted collection of extreme steering samples. This approach proved more effective than random Dagger activation, as it allowed the operator to focus data collection on challenging scenarios (sharp curves, recovery maneuvers) that are underrepresented in normal driving logs.

Offline Tangent Transformation

The application of the tangent function to steering values serves to:

• Amplify small steering differences: Low-magnitude steering commands are spread across a wider range, making the model more sensitive to subtle steering adjustments.
• Saturate extreme values: Large steering angles approach saturation, preventing the model from outputting physically unrealistic commands.
• Reduce noise sensitivity: The non-linear transformation effectively filters low-amplitude noise that could otherwise cause steering oscillations.

However, the complete effect of tangent preprocessing remains under investigation. Additional controlled experiments are required to isolate its contribution from the benefits of dataset balancing and multi-town diversity.

Category-Based Weighting Strategy

Balancing was achieved by assigning higher sampling weights to under-represented steering categories. This forced the model to learn appropriate responses across the entire steering range, preventing overfitting to straight-line driving scenarios that dominate natural driving logs.

5. Conclusions & Future Work

Proposed Work for Week 27

1. Lane discipline optimization: Implement targeted data collection focusing on lane-keeping behavior, potentially incorporating explicit lane boundary information as an auxiliary training signal.
2. Ablation study on tangent preprocessing: Conduct controlled experiments with and without tangent transformation to quantify its specific impact on performance.
3. Extended validation: Evaluate the model on additional unseen towns (Town05, Town10) to further assess generalization boundaries.
4. BehaviorMetrics integration: Implement quantitative evaluation metrics (lateral deviation, steering smoothness, turn success rate) to enable objective performance tracking.
5. Fine-tuning on extreme maneuvers: Augment the dataset with additional recovery and sharp-turn scenarios to address remaining corner cases.

Appendix: Dataset Specifications & Methodology Reference

The methodology implemented this week draws from established practices in end-to-end autonomous driving, with particular reference to the work of Jorge Rodríguez González. Key principles adopted include:

• Controlled Dagger collection: Manual triggering enables deliberate sampling of extreme steering scenarios, improving the model's ability to recover from challenging situations.
• Category-based balancing: Steering range partitioned into 25 bins, with weighting applied to ensure uniform representation across all categories.
• Multi-town training: Exposure to diverse environments (Town01, Town02, Town04, Town12) during training is essential for generalization to unseen scenarios like Town07.

All datasets were generated using CARLA simulator version 9.14, with images captured at 1920x1080 resolution and downsampled to 640x480 for training. The model architecture remains consistent with previous weeks, employing a convolutional neural network with spatial softmax for steering prediction.

Dataset availability: The balanced multi-town dataset (41,000 samples) is available for further experimentation. Additional details on preprocessing parameters and training configurations can be found in the project repository.