Enhancing Autonomous Vehicle Planning with Edge-Assisted World Models
A new approach improves trajectory predictions for Connected Autonomous Vehicles.
The planning algorithms in Autonomous Vehicles (AVs) depend on limited sensor information affected by traffic conditions and occlusions. To enhance these algorithms, a unified world model is created using data from both AVs and Road Side Units (RSUs), which aids in predicting future trajectories and improving traffic flow and collision prevention.
Introducing Conductor, an edge-based solution that generates a world model from a fixed anchor's perspective and predicts AV trajectories while adhering to a strict Age of Information (AoI) time budget. This method dynamically selects the optimal number of AVs to ensure high-quality results, particularly favoring those that detect occluded objects.
Evaluation on CAV simulation infrastructure demonstrates that this approach maintains AoI safety across various traffic scenarios, achieving fusion fidelity comparable to an Oracle and outperforming random selection methods.