kesai-labs/py123d_garage_pretrained_checkpoints

py123d_garage pretrained checkpoints

Pretrained driving policies of py123d_garage, the reference toolkit of the Alpasim E2E Challenge 2026. Each folder is one checkpoint: the weights, the training config that evaluation replays, and the sensor rigs seen in training. The README inside a folder documents its recipe, usage and finetuning.

Checkpoints

Folder	Policy	Trained on	Release
resnet34_v0.1.0	Camera-only latent TransFuser, ResNet-34 backbone	nuPlan, Physical AI AV	v0.1.0

Download

All checkpoints, into outputs/checkpoints of a py123d_garage checkout:

hf download kesai-labs/py123d_garage_pretrained_checkpoints --local-dir outputs/checkpoints

One checkpoint, by its folder:

hf download kesai-labs/py123d_garage_pretrained_checkpoints --include "<folder>/*" --local-dir outputs/checkpoints

Every evaluation entry point of py123d_garage takes the checkpoint through policy_config.evaluation_checkpoint_file and reads the config.yaml next to it. Closed-loop simulators also read a sensor rig through policy_config.evaluation_sensor_rig_file, to mount the cameras and lidar the checkpoint was trained on. Pass the rig of the dataset the simulator mimics, for example a nuPlan rig on the Alpasim nuPlan track. See the evaluation docs.

Each checkpoint is released with one py123d_garage version, listed in the table above, and is tested against that version. Newer versions aim to keep loading older checkpoints, but the replayed config.yaml follows the config schema of its release, so a schema change can break the replay. Use the code and the docs of the checkpoint's release: the folder README links to them at the release tag, and the docs linked from this page follow main. If a checkpoint fails to load on a newer version, check out its release tag.

License and citation

The checkpoints are released under Apache 2.0. The datasets they were trained on keep their own licenses: nuPlan and Physical AI AV. If you use a checkpoint, please cite py123d_garage, 123D and the datasets, see the citation notes.

@misc{py123d_garage,
  title        = {py123d_garage: end-to-end driving policies across datasets},
  author       = {KE:SAI},
  year         = {2026},
  howpublished = {\url{https://github.com/kesai-labs/py123d_garage}}
}

@article{Dauner2026ARXIV,
  title={123D: Unifying Multi-Modal Autonomous Driving Data at Scale},
  author={Dauner, Daniel and Charraut, Valentin and Berle, Bastian and Li, Tianyu and Nguyen, Long and Wang, Jiabao and Jing, Changhui and Igl, Maximilian and Caesar, Holger and Ivanovic, Boris and Geiger, Andreas and Chitta, Kashyap},
  journal={arXiv preprint arXiv:2605.08084},
  year={2026}
}