How to make autonomous cars drive safely at unsignalized intersections?

Imagine driving in a self-driving car at an intersection without traffic lights. How does it know what to do? A new study presents a smart system that combines neural networks and predictive control to make autonomous cars safer and more responsive. Unlike traditional methods, this algorithm predicts three possible trajectories for each surrounding vehicle, considering different scenarios such as yielding, not yielding, or aggressive driving. Using a graph neural network (GNN) and an attention mechanism, the system analyzes the behavior of other road users and plans the best path. Tested with real-world data, the system produces smoother and safer trajectories, reducing harsh braking and maintaining safe distances. A step forward towards more robust and comfortable autonomous driving!

The system described in the study uses a graph neural network (GNN) combined with a bidirectional gated recurrent unit (Bi-GRU) to encode agent histories and predict their future trajectories. Lane encoding is achieved through VectorNet, which provides the map context. An interaction-aware attention mechanism allows the system to focus on the most relevant agents for motion planning. Multi-head decoders generate three distinct trajectories for each agent, capturing different interaction scenarios. The model predictive control (MPC) based planner uses a bicycle model to solve a constrained optimal control problem, employing CasADi and IPOPT. The predicted trajectories are used as constraints, with the primary trajectory as a hard constraint and the alternatives as soft constraints via slack variables. This approach ensures that the autonomous vehicle can anticipate and react safely to unpredictable behaviors of other road users.