MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving
MomADv2 is a state-space memory framework for autonomous driving that solves issues with unreliable historical data during command changes. It uses a selective module (SSM-Q) to filter out inconsistent memories and retrieve only trajectory-consistent history. This, combined with a residual refiner (FM-Ref), improves long-horizon planning continuity and trajectory stability across various driving benchmarks.