Multi‐Stage Social Value Orientation‐Based Lane‐Changing Driving for Autonomous Vehicles

While intelligent driving systems are being gradually deployed on public roads, they still face numerous practical challenges. Among these, the hybrid traffic environment—comprising both autonomous vehicles and human drivers—presents significant obstacles for the development and refinement of decision‐making and planning algorithms. Human drivers exhibit diverse and often unpredictable behaviors, making it difficult to anticipate their intentions, whereas autonomous vehicles typically display uniform and conservative behavior. These differing decision‐making paradigms can significantly impact overall traffic dynamics. To address the current lack of human‐like social driving capabilities in autonomous driving algorithms and the challenge of accurately predicting other vehicles’ intentions in real time, this paper proposes a multi‐stage social value‐oriented framework for evaluating human drivers’ social driving behaviors in real time. First, environmental features are extracted from human driving datasets. During the real‐time estimation phase, inverse reinforcement learning is applied to infer the multi‐stage social value orientation of surrounding vehicles. Second, a finite‐state machine is constructed based on the estimated value orientations, enabling interactive and context‐aware planning for each stage of lane–change maneuvers. Finally, through evaluations across multiple traffic scenarios, the proposed decision‐making algorithm demonstrates its capability to generate safe, efficient, and socially compliant behaviors in mixed traffic environments.