Researchers presented a compact model for autonomous driving in CARLA

Researchers have published the results of their work on a multimodal model for controlling unmanned transport in the CARLA simulator. The system uses behavioral cloning to train on expert demonstrations, providing a closed-loop vehicle control. The model architecture includes only 1.36 million parameters. To make decisions, the system analyzes a history of five frames, combining RGB camera data, LiDAR, vehicle telemetry, and route marking points. The model predicts gas, brake, and steering angle parameters at a frequency of 20 Hz. The training dataset consisted of 236,882 windows, equivalent to 3.3 hours of real driving, collected from 448 different runs. During testing, the model demonstrated the ability to move autonomously without collisions for several hours, including passing routes that were not used during training. The developers have provided open access to the code, trained checkpoints, and the ONNX version of the model. An audit of evidence for all claimed metrics has also been published, allowing verification of the algorithm's accuracy in various road conditions, including transfer to maps with different road geometry. This work demonstrates the effectiveness of compact models in real-time control tasks. The use of the CARLA simulator allows for systematic verification of maneuvers, which is critical for the development of autonomous driving systems and reducing computational costs for on-board systems.