RESEARCH · RESEARCH · #535
Imitation Learning for Autonomous Driving in CARLA (arXiv:2609.17757v1)
The paper studies behavioral cloning in the CARLA simulator and trains a compact multimodal driving policy (1.36M parameters) using five-frame histories of RGB, LiDAR, telemetry, and lane waypoints; training used 236,882 windows (~3.3 hours of driving from 448 captures). The authors report hours-long closed-loop autonomous driving on training and held-out routes (no collisions in their runs), qualitative transfer to an unseen CARLA town, observed recovery from large deviations (not claimed as systematically evaluated), and they release code, a trained checkpoint, an ONNX model, a data sample, and an evidence audit.
KEY POINTS
- The paper studies behavioral cloning in the CARLA simulator and trains a compact multimodal driving policy (1.36M parameters) using five-frame histories of RGB, LiDAR, telemetry, and lane waypoints; training used 236,882 windows (~3.3 hours of driving from 448 captures).
- The authors report hours-long closed-loop autonomous driving on training and held-out routes (no collisions in their runs), qualitative transfer to an unseen CARLA town, observed recovery from large deviations (not claimed as systematically evaluated), and they release code, a trained checkpoint, an ONNX model, a data sample, and an evidence audit.
- Releases a compact, fully released imitation-learning driving policy and artifacts that demonstrate nontrivial closed-loop competence and cross-town transfer in CARLA, relevant for research on sample-efficient driving and reproducibility.
WHY IT MATTERS
Releases a compact, fully released imitation-learning driving policy and artifacts that demonstrate nontrivial closed-loop competence and cross-town transfer in CARLA, relevant for research on sample-efficient driving and reproducibility.