G-MARK Enhances Cooperative Driving with Knowledge Graphs

Autonomous driving systems face challenges with partial observability, where critical objects might be hidden or only visible to other connected vehicles. Existing cooperative driving methods often compress multi-agent information into abstract features, losing crucial details about which agent observed what, object visibility, and how conflicting data impacts decisions. To overcome this, G-MARK (Grounded Multi-Agent Reasoning) introduces a framework that converts cooperative object-centric observations into explicit, provenance-aware knowledge graphs (KGs). These KGs meticulously preserve object hypotheses, their source attribution, ego-versus-partner visibility, uncertainty, conflicts, spatial relationships, and relevant planning context. By deriving a shared feature representation from these KGs, G-MARK enables lightweight task heads for object reasoning, motion prediction, control selection, and trajectory forecasting. This approach leads to a 42.2% improvement in occlusion reasoning accuracy and a 13.1% reduction in control-selection error compared to state-of-the-art baselines, all while achieving comparable trajectory planning accuracy with a 25.6x smaller communication payload.