AutoE2E opens driving research weights, with non-commercial limits
AutoE2E v1.0 gives researchers a downloadable driving trajectory planner, not a ready-to-use driverless car. The Autoware Foundation announced the 79.9 million parameter model on November 11, 2026. Its source code is Apache 2.0, but the released checkpoints carry non-commercial terms. That distinction matters more than the broad “open-source” label.
The foundation’s announcement describes Autoware integration for open-loop evaluation, with real-vehicle closed-loop testing still underway. Further road trials are planned for 2027. L4 driverless capability remains a development goal, not an outcome established by this release.
What is available now
The release contains the Reactive planner. The proposed World Action and Reasoning layers are still being developed. TIER IV’s launch announcement places AutoE2E alongside other Autoware model integrations, including Alpamayo 2 Edge, METEOR, OnePlanner and LatentDrive. These are different approaches and integration stages, not interchangeable editions of one product.
AutoE2E’s source repository describes a planner that combines surround-camera imagery with vehicle-motion history and optional map information. It predicts acceleration and curvature over a 6.4-second horizon. A compact parameter count makes the research accessible to inspect, but does not establish the memory, latency or safety characteristics of an entire vehicle system.
Two checkpoints, with different evaluation roles
The Hugging Face release provides nuplan-epoch-5, trained on nuPlan, and kitscenes-epoch-5, further fine-tuned on KITScenes-Multimodal. These are PyTorch training checkpoints with optimizer and training state, not a packaged driving application. The loading instructions pin a specific training-source revision. Follow that revision rather than assuming the latest repository code reproduces the release.
For a research evaluation, record the checkpoint, source revision, camera configuration and dataset split together. Changing those at the same time makes it difficult to tell whether a difference came from the model or the test setup. ByteForward has not run these checkpoints or independently reproduced their results.
What the benchmark does and does not show
Autoware reports a three-second average displacement error of 0.744 meters for the fine-tuned model on KITScenes validation, second among eight compared configurations. This is an open-loop measurement: a predicted path is compared with a previously recorded human path. It does not measure how the system reacts when its own decisions change the scene.
The repository’s comparison also uses different sample counts and inputs across models. AutoE2E’s quoted result uses 202 keyframes plus map and route information. Treat the ranking as a documented experiment, not a universal leaderboard for road performance. A useful replication needs the same input contract, horizon and sample selection before the numbers can be compared meaningfully.
The model card adds that validation routes are reconstructed from future logged trajectories. It also notes that confidence intervals and per-scenario results are not yet published. Its intended-use section limits v1.0 to research and explicitly says it must not control vehicles on public roads.
Apache code does not mean commercial weights
The source repository separately identifies Apache 2.0 code and CC BY-NC-SA 4.0 checkpoints. The model card says commercial checkpoint use is not covered and requires upstream permissions, including a Motional commercial license. Camera imagery and dataset files have their own terms. Check the exact artifact before redistributing it or building a commercial workflow around it.
The practical opportunity is a reproducible planning experiment, with a clear boundary between downloadable research and deployment. Our PAN 2 world-model guide explores the related problem from simulation: a plausible predicted outcome is not the same as a reliable action. AutoE2E offers another inspectable component, not a shortcut around system validation.
