This study presents KEPT, an AI system that helps self-driving cars predict their own short-term path more safely by combining video understanding with a memory of similar past scenes. Tested on the public nuScenes benchmark, KEPT cuts prediction errors and potential collisions compared with existing planning methods, while using a fast, lightweight retrieval module that is practical for real-time driving.

To address the growing conflict between personalized mobility analysis and data privacy, researchers have developed IPC-FM, a novel federated meta-learning framework. This approach enables accurate travel behavior prediction without centralizing sensitive user data. By integrating interpretable neural networks with rapid model adaptation, IPC-FM provides a customizable solution that significantly outperforms current state-of-the-art methods, ensuring individual mobility needs are met securely and transparently.

Researchers at Beihang University, China, introduce a new task setting: latency-aware trajectory prediction for autonomous driving, which explicitly accounts for the latency issue and transforms it from a hindrance into an opportunity for enhanced performance.

How can autonomous vehicles continuously learn new traffic scenarios without forgetting previously learned ones? Researchers from Tsinghua University have proposed a dynamically expandable learning framework for interactive trajectory prediction. The method enables models to adapt to evolving traffic environments while preserving performance on earlier scenarios. Experiments on real-world datasets show that the approach effectively mitigates catastrophic forgetting, especially for safety-critical driving cases.

Mandatory lane changes at intersections often lead to intricate conflicts and traffic oscillations.  The advent of connected and autonomous vehicles (CAVs) is expected to mitigate these disruptions by coordinating acceleration and lane-change behaviors. Addressing this, researchers developed SS-MA-PPO, a novel Multi-Agent Reinforcement Learning (MARL) framework that assists CAVs in coordinating these critical decisions. Evaluated against a real-world dataset from Langfang, this method significantly improves traffic efficiency compared with traditional models and other Multi-Agent Reinforcement Learning baselines.