Autoware Foundation releases AutoE2E, an open-source driving model

The foundation calls v1.0 the Reactive model and says it runs at 10 Hz: ten times a second it fuses six camera views into a bird's-eye view of the road and plans the next 6.4 seconds of acceleration and steering. The foundation calls this the layer that drives[1]. The repository describes a GRU planner that turns that bird's-eye scene, an optional map raster and the vehicle's recent motion into 6.4 seconds of acceleration and curvature[3]. The model card states the same output as 64 steps of longitudinal acceleration and curvature at 10 Hz, integrated into a trajectory in the vehicle's frame[4].

Two slower layers, both at 1 Hz, are still in development. One is a JEPA world model that predicts how the scene will evolve without rendering pixels. The other is a language-free reasoning layer that works out why the car should act, such as easing off for a pedestrian about to cross, without generating a word[1]. The model card says both branches are disabled in the v1.0 weights, and the foundation says every published result comes from the Reactive model alone[1][4].

The foundation says a camera-only setup is for hands-free driving on highways and city streets, and that the output can be paired with LiDAR and radar safety layers for fully driverless applications[1]. TIER IV says the model can cover configurations from cameras alone through to LiDAR and radar[2].

On KITScenes-Multimodal validation, the foundation reports an average displacement error of 0.744 metres over three seconds for the fine-tuned model, ranking second among eight open-weight model configurations[1]. The repository and the model card put that figure on 202 keyframes, with an error of 1.980 metres at the three-second point. In the foundation's table, only Qwen Drive 1.0 is ahead on those keyframes, at 0.565 metres average displacement and 1.397 metres at three seconds[3][4].

The foundation says the comparison is open-loop: each model predicts the next three seconds from logged camera frames, and the score is the distance from the path the human driver took[1][3]. AutoE2E also receives a map raster and a route. The model card says that route is an oracle rebuilt from the logged future trajectory, so the figure measures route-conditioned driving. It says the validation ground truth comes from a trajectory field made with an older coordinate conversion, that rescoring against the official pose file is still open, and that closed-loop behaviour is not measured in these results[4].

A nuPlan-trained checkpoint, evaluated zero-shot, is listed at 0.915 metres average displacement over three seconds on 11,035 validation samples, a larger set than those 202 keyframes[3][4]. On the official KITScenes test split, cameras and ego-motion only, the fine-tuned checkpoint is listed at 0.826 metres average displacement over three seconds on 23,690 samples. The model card says neither checkpoint saw those scenes in training[4].

Source code is on GitHub under the Apache 2.0 licence, which the foundation says allows research and commercial use[1][3]. The v1.0 checkpoints are on Hugging Face[1][4]. One, nuplan-epoch-5, is trained on nuPlan v1.1 and is the baseline for a new city or a fine-tune. The other, kitscenes-epoch-5, is that checkpoint fine-tuned on KITScenes-Multimodal for Karlsruhe and the KITScenes camera rig[4].

The model card puts the checkpoint files under CC BY-NC-SA 4.0. They may be used, fine-tuned and shared for non-commercial work, with attribution, and shared adaptations stay under the same licence. Commercial use needs permission from the upstream rights holders, including a commercial licence from Motional covering nuPlan and nuScenes material used to build the weights[3][4]. The card says v1.0 is a research model for open-loop evaluation, visualisation, benchmarking and further training. It says the release is not a certified driving system and must not control a vehicle on public roads[4].

The Reactive model is integrated with Autoware for open-loop evaluation[1][2]. The foundation says real-vehicle closed-loop testing is underway, with further on-road trials planned for 2027, alongside continued development and automotive system-on-chip optimisation. It says those steps support eventual commercialisation, that L4 capability remains a development goal, and that deployment performance remains subject to evaluation on each platform[1]. TIER IV says robotaxi commercialisation is under consideration while the world model and the reasoning model are still being developed[2].

That 0.744 metre result is the foundation's open-loop score on 202 validation keyframes, with a map and a route supplied. The code is under Apache 2.0. The checkpoints carry a non-commercial licence, and the model card says they must not control a vehicle on public roads.