Behavior Cloning for Mini Autonomous Car Path Following

TL;DR

The work investigates behavior cloning for route following on mini autonomous cars using a CNN that maps front-camera input to steering and throttle commands trained from human driving data. An ablation study compares three CNN configurations on a 3×3 meter track, evaluating performance via lap times and trajectory deviations. Results show a trade-off between central-line accuracy and stability, with the original architecture delivering best tracking at the cost of more offs, while altering the network depth changes both stability and speed. The findings demonstrate the viability of behavior cloning for scalable autonomous driving prototypes and offer guidance on CNN design for edge-device deployment, with potential extensions to full-sized vehicles and reinforcement-learning-based parameter tuning.

Abstract

This article presents the implementation and evaluation of a behavior cloning approach for route following with autonomous cars. Behavior cloning is a machine-learning technique in which a neural network is trained to mimic the driving behavior of a human operator. Using camera data that captures the environment and the vehicle's movement, the neural network learns to predict the control actions necessary to follow a predetermined route. Mini-autonomous cars, which provide a good benchmark for use, are employed as a testing platform. This approach simplifies the control system by directly mapping the driver's movements to the control outputs, avoiding the need for complex algorithms. We performed an evaluation in a 13-meter sizer route, where our vehicle was evaluated. The results show that behavior cloning allows for a smooth and precise route, allowing it to be a full-sized vehicle and enabling an effective transition from small-scale experiments to real-world implementations.

Figures (5)

• Figure 1: Specifications of the Autonomous Car
• Figure 2: Autonomous car in evaluation environment
• Figure 3: Network Structure
• Figure 4: Loss of the CNN models used for evaluation.
• Figure 5: Track Scenario Used to Validate Our Vehicle.