AutoE2E checkpoints
Experimental contributor checkpoint notice
These artifacts are not an official Autoware release. They are experimental checkpoints published for contributors to reproduce evaluation results, inspect model behavior, and continue research. They are not release-qualified or safety-qualified models.
This repository contains four trajectory-planning checkpoints from the AutoE2E project. These are the exact PyTorch training checkpoints registered in MLflow, including model, optimizer, scheduler, configuration, metrics, and training state.
Checkpoints
| Path | Training stage | Epoch | MLflow version | SHA-256 |
|---|---|---|---|---|
models/nuplan-epoch-4/checkpoint.pt |
nuPlan trajectory and route | 4 | 64 | ed00e072471aae51b4cf48fa54c3fd9c03c61e0fd6e2978829cd9068dfd6da68 |
models/nuplan-epoch-5/checkpoint.pt |
nuPlan trajectory and route | 5 | 65 | ca8b43d7a777d6fd9195bb253d1452f3cd645f7aab3bd0e36b1f74ffb31df29b |
models/kitscenes-epoch-5/checkpoint.pt |
KITScenes fine-tuning | 5 | 66 | 120a21639d9767d512eea645be0a83b01476c9e0366a8b957ac51b7ead1bd762 |
models/kitscenes-epoch-7/checkpoint.pt |
KITScenes fine-tuning | 7 | 67 | a1e6b1621018740485b3e1e43dda23713d913667eb39a65b77fcbf895dcbcefd |
All checkpoints use the bevformer_v2_t8_split_navigation_v5 configuration. The camera BEV starts from the official BEVFormer V2 R50 T8 initialization and is frozen during trajectory stages. World Model and Reasoning branches are disabled.
KITScenes Val v3.5
KITScenes Val v3.5 contains 11,035 samples. Evaluation uses Camera, HD Map, and an oracle Route reconstructed retrospectively from the logged future trajectory.
Each horizon value is ADE / FDE in meters.
| MLflow version | Model | 1 s | 2 s | 3 s | 5 s | Lateral error | Longitudinal error |
|---|---|---|---|---|---|---|---|
| 64 | nuPlan Epoch 4 | 0.2030 / 0.4066 | 0.5018 / 1.1935 | 0.9503 / 2.4508 | 2.2919 / 6.2272 | 1.2082 | 1.6448 |
| 65 | nuPlan Epoch 5 | 0.1951 / 0.3779 | 0.4723 / 1.1331 | 0.9147 / 2.4226 | 2.2951 / 6.3885 | 1.1911 | 1.6633 |
| 66 | KITScenes Epoch 5 | 0.1472 / 0.2850 | 0.3729 / 0.9239 | 0.7449 / 2.0196 | 1.9405 / 5.5645 | 0.9844 | 1.4025 |
| 67 | KITScenes Epoch 7 | 0.1347 / 0.2831 | 0.3886 / 1.0101 | 0.8035 / 2.2127 | 2.0952 / 5.9509 | 1.0220 | 1.5659 |
KITScenes Test v1.0
KITScenes Test v1.0 contains 23,690 samples. This evaluation is Camera-only and uses a different scene population from KITScenes Val.
Each horizon value is ADE / FDE in meters.
| MLflow version | Model | 1 s | 2 s | 3 s | 5 s | Lateral error | Longitudinal error |
|---|---|---|---|---|---|---|---|
| 64 | nuPlan Epoch 4 | 0.1916 / 0.3899 | 0.4817 / 1.1514 | 0.9181 / 2.3895 | 2.2840 / 6.4266 | 1.2167 | 1.6287 |
| 65 | nuPlan Epoch 5 | 0.1602 / 0.3134 | 0.3979 / 0.9727 | 0.7924 / 2.1621 | 2.1120 / 6.1751 | 1.0584 | 1.5496 |
| 66 | KITScenes Epoch 5 | 0.1463 / 0.3107 | 0.4121 / 1.0453 | 0.8263 / 2.2199 | 2.1042 / 5.9422 | 1.2044 | 1.4051 |
| 67 | KITScenes Epoch 7 | 0.1680 / 0.3754 | 0.4923 / 1.2512 | 0.9831 / 2.6242 | 2.4699 / 6.8964 | 1.4730 | 1.6598 |
Evaluation scope
The external evaluation is a deterministic replay of published control overlays. It does not re-run the model. Val and Test are different populations, so their difference is not evidence of a causal benefit from map or route inputs. The complete index and eight reports are under evaluations/overlay-replay-v1/.
The KITScenes Val route contains oracle maneuver information because it is reconstructed post hoc from the logged future trajectory. Val results must not be interpreted as online route-planning performance.
Loading
import torch
checkpoint = torch.load("checkpoint.pt", map_location="cpu", weights_only=True)
model.load_state_dict(checkpoint["model_state_dict"], strict=True)
Use an AutoE2E source revision compatible with bevformer_v2_t8_split_navigation_v5. Each model directory contains metadata.json. SHA256SUMS verifies all published files.
Intended use and limitations
These checkpoints are experimental research artifacts for contributor evaluation and continued training. They are not an official Autoware release and are not validated for safety-critical deployment. Distribution shift remains relevant.
License notice
The AutoE2E source code is Apache-2.0. These checkpoint files include parameters initialized from the official BEVFormer V2 R50 T8 checkpoint. The upstream weight license was recorded as NOASSERTION, and the pretrained component was trained on nuScenes. Users are responsible for complying with the terms of the upstream weights and datasets. No broader checkpoint license is asserted here.