LaksNet: An End-to-End Deep Learning Model for Self-driving Cars in Udacity Simulator

Lakshmikar R. Polamreddy, Youshan ZhangLecture Notes in Networks and Systems 2023 · DOI · OpenAlex

Abstract

This paper introduces LaksNet, a lightweight convolutional neural network (CNN) designed for end-to-end autonomous vehicle control within the Udacity simulator. While industry trends often favor massive, pre-trained architectures designed for image classification, LaksNet is optimized specifically for the regression task of steering angle prediction. By prioritizing a streamlined architecture, LaksNet achieves superior autonomous drive times with approximately half the parameters of the industry-standard NVIDIA model. Our results demonstrate that model complexity does not inherently correlate with performance in dynamic environments, and that task-specific architectural design is critical for generalization.

1. Introduction

The "holy grail" of autonomous driving research is end-to-end learning: the capacity for a vehicle to map raw sensory input, such as camera pixels, directly to control commands like steering angles. For years, the NVIDIA model has served as the primary benchmark for this task. However, a common pitfall in current research is the uncritical adoption of massive architectures—such as ResNet or Xception—originally designed for the ImageNet classification challenge.

Driving is fundamentally a regression problem, requiring the prediction of continuous numerical values, whereas ImageNet models are optimized for classification (identifying discrete categories). This study investigates whether these heavy, classification-centric models are truly suitable for driving and proposes LaksNet as a leaner, more robust alternative that avoids overfitting to static training data.

2. The LaksNet Architecture

LaksNet was developed through extensive hyperparameter experimentation to identify the optimal balance between feature extraction and computational efficiency. The architecture is characterized by its streamlined design:

  • Convolutional Layers: The model utilizes four convolutional layers. The first three employ $3 \times 3$ kernels to capture fine-grained local details, while the final layer utilizes a $5 \times 5$ kernel to extract broader spatial features.
  • Regularization and Dimensionality Reduction: Max-pooling and dropout layers are integrated after every convolutional block to reduce spatial dimensionality and mitigate the risk of overfitting.
  • Fully Connected Layers: The feature maps are flattened into two dense layers, culminating in a single output node representing the predicted steering angle.

LaksNet is remarkably efficient, utilizing only 274,017 parameters. In comparison, the standard NVIDIA architecture contains 559,419 parameters—more than double the footprint of LaksNet.

3. Experimental Methodology

3.1 Dataset and Training

The model was trained and validated using the Udacity self-driving car simulator. The training dataset comprised 130,426 images, consisting of 33,096 images generated by the authors and 97,330 images sourced from open-source datasets. Input images were processed at a resolution of $70 \times 320 \times 3$.

To ensure robustness, the authors applied data augmentation techniques, including random rotations and cropping. The model was optimized using the Adam optimizer over 50 epochs, utilizing Mean Squared Error (MSE) as the loss function.

3.2 Evaluation Metrics

Models were evaluated based on their ability to maintain autonomous control at a constant speed of 10 mph. The primary metric was Run Time (Seconds)—the duration the vehicle remained on the track without manual intervention or going off-course.

4. Results

LaksNet significantly outperformed both the industry-standard NVIDIA model and several high-capacity ImageNet-based architectures.

Model Parameters Run Time (Seconds)
LaksNet 274,017 150
NVIDIA 559,419 120
AlexNet ~60 Million 50
GoogleNet ~7 Million 50
ResNet101 ~44 Million 14

LaksNet achieved a 25% increase in autonomous duration over the NVIDIA model. Notably, the massive models (ResNet101, NasnetLarge) performed poorly, often failing within 3 to 14 seconds of operation.

5. Discussion

5.1 Complexity vs. Performance

The failure of deep models like ResNet101 suggests they are "too smart" for the specific task of simulated steering. Their depth allows them to extract highly abstract features that may be irrelevant to lane-keeping, leading to poor performance in dynamic environments. LaksNet’s shallower structure forces the network to focus on the most salient visual cues for steering.

5.2 The Generalization Paradox

An interesting finding was that LaksNet exhibited a higher MSE on static test images compared to the NVIDIA model. However, LaksNet performed better in the live simulator. This suggests that the NVIDIA model may be slightly overfitted to specific frames in the training set, while LaksNet learned more generalized "rules of the road," allowing it to recover more effectively when the vehicle drifted from the center of the lane.

6. Conclusion and Future Work

LaksNet demonstrates that for end-to-end driving in simulated environments, a smaller, task-specific regression model outperforms larger, classification-oriented networks. Future work will focus on expanding the model to include throttle control (acceleration and braking) and addressing the "sim-to-real" gap to determine if this lightweight architecture remains robust against the noise and unpredictability of real-world road conditions.

Summary updated

BibTeX
@inproceedings{polamreddy2023laksnet,
  author = {Lakshmikar R. Polamreddy and Youshan Zhang},
  title = {{LaksNet: An End-to-End Deep Learning Model for Self-driving Cars in Udacity Simulator}},
  booktitle = {Lecture Notes in Networks and Systems},
  year = {2023},
  doi = {10.1007/978-3-031-47448-4\_1},
}

Loading related content...

Figure