Towards Selection and Transition Between Behavior-Based Neural Networks for Automated Driving
This addresses safety and interpretability concerns in autonomous driving systems, though it is an incremental improvement over existing modular approaches.
The paper tackles the problem of complex and opaque end-to-end neural networks in autonomous driving by introducing a Behavior Selector that uses multiple smaller networks for different tasks, enabling dynamic real-time selection and smooth transitions to improve stability and safety, with effectiveness demonstrated in AirSim simulations.
Autonomous driving technology is progressing rapidly, largely due to complex End To End systems based on deep neural networks. While these systems are effective, their complexity can make it difficult to understand their behavior, raising safety concerns. This paper presents a new solution a Behavior Selector that uses multiple smaller artificial neural networks (ANNs) to manage different driving tasks, such as lane following and turning. Rather than relying on a single large network, which can be burdensome, require extensive training data, and is hard to understand, the developed approach allows the system to dynamically select the appropriate neural network for each specific behavior (e.g., turns) in real time. We focus on ensuring smooth transitions between behaviors while considering the vehicles current speed and orientation to improve stability and safety. The proposed system has been tested using the AirSim simulation environment, demonstrating its effectiveness.