A Brain-Inspired Perception-Decision Driving Model Based on Neural Pathway Anatomical Alignment

TL;DR

This work proposes a novel brain-inspired driving (BID) framework that harnesses brain-inspired perception technology to achieve more efficient and robust environmental perception and employs brain-inspired decision-making techniques to facilitate intelligent decision-making.

Abstract

In the realm of autonomous driving, conventional approaches for vehicle perception and decision-making primarily rely on sensor input and rule-based algorithms. However, these methodologies often suffer from lack of interpretability and robustness, particularly in intricate traffic scenarios. To tackle this challenge, we propose a novel brain-inspired driving (BID) framework. Diverging from traditional methods, our approach harnesses brain-inspired perception technology to achieve more efficient and robust environmental perception. Additionally, it employs brain-inspired decision-making techniques to facilitate intelligent decision-making. The experimental results show that the performance has been significantly improved across various autonomous driving tasks and achieved the end-to-end autopilot successfully. This contribution not only advances interpretability and robustness but also offers fancy insights and methodologies for further advancing autonomous driving technology.

Figures (6)

• Figure 1: Perception-decision network diagram based on neural pathway anatomical alignment
• Figure 2: The network architecture of the BID agent is primarily composed of the perception network and the decision network. The perception module includes both the dorsal and ventral pathways, which allow the system to understand and process its surroundings. The decision network consists of five modules that represent different regions of the prefrontal cortex: the medial prefrontal cortex (MPC), orbital prefrontal cortex (OPC), caudal prefrontal cortex (CPC), dorsal prefrontal cortex (DPC), and ventral prefrontal cortex (VPC). Specifically, the input to the system is an RGB image $f_{t}$, which represents the current visual scene. The output is the action $A_t$, which controls the ego-vehicle's maneuvering directly. The action $A_t$ includes the steering angle $A_t^s$ and acceleration $A_t^a$ of the vehicle, making BID a brain-inspired and end to end autonomous driving model. Besides, in the construction process of training dataset, the driving actions and input frames are paired.
• Figure 3: Top: When the ego-vehicle is driving on the road, even though the triffic light is green, the vehicle in front has stopped, but the ego-vehicle still brakes. Bottom: When the vehicle is moving and encounters a traffic jam in front, it will automaitcally brake and stop. The brake command in the figure is displayed in red.
• Figure 4: BID architecture analysis. We analyse four main factors including the initial step size in ventral pathway, the number of recurrent step in dorsal pathway, the position feature dimension in VPC and the vector embedding dimension in DPC. Each row presents how SR and TL score change based on BID when the hyperparameter changes.
• Figure 5: Driving performance and training progress of BID. All BID agents are evaluated in LBCRoutes after 30 peoch. Top figure: The ground truth and model prediction for steering angles are shown in green and blue, respectively. Bottom figure: The errors in training progress are displayed every 5 epochs, including the mean absolute error (MAE) for steering and acceleration. While other metrics are tested with single random number, we describe the outcome as the averaged value over 5 test random number. For evaluation, the offical benchmark is employed.