Integrating social neuroscience into human-machine mutual behavioral understanding for autonomous driving

摘要

Autonomous vehicles (AVs) are advertised to free human drivers, providing a safer and more efficient transport mode. After decades of extensive investment and invention, various types of AVs have been unveiled, but they are still restricted to limited application scenarios because of potential safety concerns. Despite rare sensing or detection failures from corner cases, one of the significant concerns primarily questions whether AVs would interact appropriately with surrounding human-driven vehicles on public roads. Particularly, the lack of approaches to human-like mutual driving understanding challenges the driving safety situation of AVs and human drivers. That is, human informal driving rules and implicit driving interactions cannot be understood explicitly by AVs, and human drivers can hardly accommodate the stilted or inconsistent driving behaviors generated by AVs because of distinctive “driving behavior understanding” mechanisms. Due to the current small market penetration rate of AVs, they will inevitably share roads with human-driven vehicles for a long term. In other words, interactions between human-driven vehicles and AVs may challenge road safety for a prolonged period. Moreover, the opacity of AV decision-making algorithms brings psychological roadblocks to human drivers’ trust, which prevents public acceptance and affects the adoption of AVs. Therefore, barriers to mutual driving understanding underline the urgent need for investigation. Because driving interaction could be formulated as a cooperative task, both vehicle types require mutual understanding and cooperative road sharing. In this Editorial, we point out the current research bottleneck in AV driving understanding and gain insights into the integration with state-of-the-art social neuroscience research, such as interbrain synchrony or social neuromorphic computing, to achieve human-machine mutual driving understanding. Recent research advances confirm that state-of-the-art AV research is dominated by various data-driven models, such as deep learning models and their variants. Under this technology trend, most AV modules have been facilitated from driving scene understanding tasks to end-to-end autonomous route planning frameworks. Admittedly, AVs informed by these data-driven models function well under restricted laboratory environments or free-flow traffic conditions. However, the black-box nature of these data-driven models brings shortcomings and research bottlenecks for AVs to achieve mutual or even unidirectional driving understanding. Generally, because data-driven models learn from data rather than explicit rules or knowledge, they are believed to have inherent reproducibility and interpretability bottlenecks. A plausible standpoint is that data-driven neural networks use node gradients and weights across different layers to fit all kinds of correlations between input data and training targets rather than proposing causal reasoning or mechanism-level analysis. Because proposed fittings in these black boxes lack physical meaning, the model may put out outliers, and confounding model variables are difficult to troubleshoot, which is unacceptable in safety-critical applications like driving. The bottlenecks of data-driven models are magnified when we focus on their employment in AV’s driving understanding tasks. From the model’s perspective, the “state space” of driving behaviors and interactions is staggeringly vast, requiring massive recordings of human driving demonstrations. From the task’s perspective, real-time interaction characteristics of driving tasks make data-driven model deployment incapable and error-prone. Execution of misleading or erratic driving behavior generated by AVs is inclined toward traffic accidents or even fatalities, which threatens the traffic safety situation. Moreover, better model performance does not equal higher reproducibility or interpretability because the relationships discovered from driving data do not nece