Recently, Professor Zhu Meixin from our school as the first corresponding author, published a research paper titled "Empowering safer socially sensitive autonomous vehicles using human-plausible cognitive encoding" in the Proceedings of the National Academy of Sciences (PNAS). The study proposes an innovative framework based on social attention and human-plausible cognitive encoding, aiming to enhance the ethical decision-making capabilities of autonomous vehicles (AVs) in complex traffic environments and achieve safer and more equitable risk management.
With the rapid development of autonomous driving technology, the scenario of autonomous vehicles sharing the road with humans is becoming increasingly common. However, current autonomous vehicles (AVs) lack social sensitivity in ethical decision-making and struggle to simultaneously balance the differentiated protection needs of different road users and their collective impact. For example, when facing pedestrians, cyclists, and trucks, how can the risks of different road users be balanced? When multiple road users appear simultaneously (e.g., a pedestrian suddenly dashing out + a car overtaking from the right), how can AVs consider the risks of each individual while also ensuring overall safety? Traditional solutions either treat everyone equally or fail to balance one aspect with another.To address this issue, inspired by neuroscience and cognitive behavioral science, the research team has proposed a human cognitive encoding scheme (SR active inference). By encoding the risks of multiple road users into behavioral beliefs, AVs can achieve social sensitivity in ethical decision-making, much like human drivers.
The study introduces the human cognitive encoding mechanism into the ethical decision-making of autonomous driving for the first time, providing a practical technological pathway to resolve ethical dilemmas such as the "trolley problem." The research team emphasized that the weights of social attention can be flexibly adjusted according to cultural differences, and the fairness and transparency of the framework can be further optimized through public consensus in the future. The study also pointed out that this solution can be extended to more scenarios, such as interactive information and customized passenger needs, laying the foundation for the integration of autonomous driving into the social traffic environment.