A new research paper shows a surprisingly small driving model can hold its own in closed-loop simulation, not just in offline tests.
Researchers trained a 1.36 million parameter multimodal policy in the CARLA simulator using behavioral cloning, feeding it five-frame histories of RGB images, LiDAR, vehicle telemetry, and lane waypoints to predict throttle, brake, and steering at 20 Hz. The training set came from 236,882 windows, roughly 3.3 hours of driving pulled from 448 captures, collected through a staged process that culminated in a systematic route-generation method enumerating spawn points and verified autopilot maneuvers. In testing, the policy drove for hours on both training and held-out routes without collisions. It also showed qualitative transfer to an unseen CARLA town with different road geometry, and recovered from some large trajectory deviations, though the team stops short of claiming reliable recovery without controlled evaluation.
The real story here is the size-to-competence ratio. Most closed-loop driving demos come from much larger models or reinforcement learning setups that need extensive online interaction. This one is trained purely offline on a modest dataset and still holds together in closed-loop conditions, where every action changes what the policy sees next - the exact feedback loop that usually breaks behavioral cloning. That gap between offline accuracy and closed-loop stability is where a lot of imitation-learning driving research quietly falls apart.
The team released the code, checkpoint, an ONNX version, a data sample, and an audit trail for their claims - a level of transparency that's still rare in autonomy research and worth noting on its own. Still, this is simulation only, and CARLA's clean lane markings and forgiving physics are a long way from a rainy intersection with an unpredictable cyclist.