AutoE2E goes open source under Apache-2.0: a language-free driving model with 79.9M parameters, only 1.45M of them trainable
Synopsis
The Autoware Foundation's Robotaxi Working Group has released AutoE2E v1.0, a deterministic end-to-end driving model with no language component that fuses six camera views, HD-map rasters and a selected-route raster into a bird's-eye-view representation and outputs a trajectory directly. The model has about 79.9M parameters, of which only about 1.45M are trainable in the evaluated configuration, the rest belonging to a frozen camera BEV backbone. On the KITScenes validation set, the fine-tuned checkpoint reaches an average displacement error of 0.745 m and a final displacement error of 2.020 m at a three-second horizon; that evaluation is open-loop replay, and the model has not yet demonstrated validated closed-loop driving.
Interpretation
The open-source driving model AutoE2E v1.0 combines camera, HD-map and route inputs into one bird's-eye-view representation, from which a deterministic planner emits a future trajectory without sampling multiple candidates at inference. The representative systems on this route have been vision-language-action planners in the 4B to 10B parameter range, relying on teacher-student distillation to fit an on-vehicle inference budget. On 11,035 samples from 117 KITScenes validation scenes, the fine-tuned checkpoint reaches an average displacement error of 0.745 m and a final displacement error of 2.020 m at three seconds, and 1.94 m and 5.56 m at five seconds.
The model attaches a new navigation pathway to a frozen temporal camera bird's-eye-view backbone; the pathway encodes the HD-map raster and the route raster separately and injects them into the camera representation through a gated residual and deformable cross-attention. Map and route have separate encoders, and the route contribution is controlled by a zero-initialized per-channel gate whose weight is 0.5 at the start of training; the fusion's output projection and final linear layer are zero-initialized, so at the start of training the fusion leaves the pretrained camera backbone's representation unchanged. Of 79,906,522 total parameters, 97.98% belong to the frozen camera pathway, and only 1,448,582 are updated during trajectory training, spread across the GRU planner, the navigation fusion, the map and route encoders, and the route-reconstruction head.
Perspective
The demonstrated scope is open-loop trajectory replay on the KITScenes validation and test splits, with the primary evaluation under the condition where camera, HD map and route are all available, and the test split serving only as a camera-only robustness track because its maps are not released. The intended readers are researchers and engineering teams who want to reuse or audit an end-to-end driving baseline; the real-vehicle portion is preliminary research on the Technical University of Munich's EDGAR platform.
The evidence does not support a causal claim that the HD map and route improve planning accuracy, because validation and test are different scene populations and no same-scene camera-only comparison exists. The external replay covers only 5 s, leaving about 22% of the 6.4 s output unevaluated; the route is fitted after the fact to already-driven lanes, so robustness to route errors is not measured. In a first closed-loop integration with NVIDIA's AlpaSim, the baseline checkpoint completed five NuRec scenes without collision but left the road in four of five within 2.5 to 4.1 s of takeover.
