Compact 1.36M Parameter Multimodal Policy for Autonomous Driving
September 17, 2026
A 1.36 million parameter multimodal policy trained on 236,882 windows of CARLA simulator data enables closed-loop autonomous driving. The model processes five-frame histories of RGB, LiDAR, telemetry, and lane waypoints to predict throttle, brake, and steering at 20 Hz.
HOW THIS AFFECTS YOU
●
builderYou can leverage extremely low-parameter multimodal architectures for real-time robotics applications.
●
researcherThis demonstrates the efficacy of compact multimodal policies in high-frequency closed-loop control tasks.