DecisionPerceiver: New AI boosts autonomous driving lane changes with attention

Researchers from multiple institutions have introduced DecisionPerceiver, a novel attention-based architecture for high-level decision making in autonomous driving. The system handles lane changes, speed control, and intersection negotiations by explicitly modeling interactions among vehicles and static elements. Unlike prior DeepSet approaches that ignore traffic interaction dynamics, DecisionPerceiver leverages an attention mechanism inspired by Perceiver IO to project dynamic agent features into a fixed-size latent space. This design drastically reduces quadratic memory and computation costs typically associated with full attention, enabling the network to scale gracefully as the number of road users increases.

Extensive tests across three distinct driving scenarios—each demanding different levels of interaction awareness—show consistent performance gains over state-of-the-art shared-encoder methods. The architecture also accommodates a finer discretization of the action set, which amplifies the benefits of interaction modeling. Notably, DecisionPerceiver maintains robust generalization when navigation objectives change, and its scalability is validated in simulations with up to dozens of vehicles. The paper, submitted to the 2026 IEEE Intelligent Transportation Systems Conference, positions DecisionPerceiver as a practical step toward reliable, learning-based autonomous driving.