UniLION: Towards Unified Autonomous Driving Model with Linear Group RNNs

Zhe Liu ^{1, 2}, Jinghua Hou ¹, Xiaoqing Ye ³, Jingdong Wang ³, Hengshuang Zhao ^2,✉, Xiang Bai ^1,✉
1 Huazhong University of Science and Technology, ² The University of Hong Kong, ³ Baidu Inc.
✉ Corresponding author.

🔥 Highlights

• Unified Heterogeneous Inputs. UniLION integrates multi-view images, LiDAR point clouds, and temporal information into a unified 3D backbone through direct token concatenation without hand-crafted fusion modules. 💪

• Unified Model. UniLION enables parameter sharing across different input formats. Specifically, once trained with multi-modal temporal data, the same UniLION model can be directly deployed across different sensor configurations and temporal settings (e.g., LiDAR-only, temporal LiDAR, or multi-modal fusion) without retraining. 💪

• Unified Representation. UniLION highly compresses heterogeneous multi-modal and temporal information into a compact BEV feature representation that serves as a shared strong representation for autonomous driving. 💪

• Strong performance. UniLION achieves competitive and SOTA performance across comprehensive autonomous driving tasks including 3D perception, motion prediction, and planning. 💪

News

• 2025.12.15: DrivePI paper released. 🔥
• 2025.12.15: GenieDrive (Physics-Aware Driving World Model) paper released. 🔥
• 2025.06.16: Our new work about Transformer-Mamba architecture HybridTM have been accepted by IROS 2025 as Oral presentation. 🎉
• 2024.09.26: LION has been accepted by NeurIPS 2024. 🎉
• 2024.07.25: LION paper released. 🔥
• 2024.07.02: Our new works OPEN and SEED have been accepted by ECCV 2024. 🎉

Results

• nuScenes Validation Set

Model	Modality	NDS	mAP	AMOTA	mIoU	RayIoU	minADE (Car/Ped.)	L2	Col.	Config	Checkpoint
UniLION	L	72.3	67.5	72.6	71.7	46.8	-	-	-	-	-
UniLION	LT	73.0	68.9	73.3	72.4	49.6	0.58 / 0.39	0.60	0.27	-	-
UniLION	LC	74.9	72.2	76.2	72.3	50.8	-	-	-	config	-
UniLION	LCT	75.4	73.2	76.5	73.3	51.3	0.57 / 0.37	0.65	0.18	config	model

• 3D Object Detection

Model	Modality	NDS	mAP
UniLION	L	72.3	67.5
UniLION	LT	73.0	68.9
UniLION	LC	74.9	72.2
UniLION	LCT	75.4	73.2

• Multi-object Tracking

Model	Modality	AMOTA	AMOTP	IDS
UniLION	L	72.6	0.542	510
UniLION	LT	73.3	0.515	537
UniLION	LC	76.2	0.499	711
UniLION	LCT	76.5	0.477	613

• BEV Map Segmentation

Model	Modality	mIoU
UniLION	L	71.7
UniLION	LT	72.4
UniLION	LC	72.3
UniLION	LCT	73.3

• 3D Occupancy Prediction

Model	Modality	RayIoU
UniLION	L	46.8
UniLION	LT	49.6
UniLION	LC	50.8
UniLION	LCT	51.3

• Motion Prediction

Model	Modality	minADE (Car)	minADE (Ped)	EPA
UniLION	LT	0.58	0.39	0.647
UniLION	LCT	0.57	0.37	0.678

• Planning

Model	Modality	L2 (1s)	L2 (2s)	L2 (3s)	L2 (avg)	Col. (1s)	Col. (2s)	Col. (3s)	Col. (avg)
UniLION	LT	0.35	0.67	1.09	0.70	0.01	0.20	0.60	0.27
UniLION	LCT	0.33	0.62	0.99	0.65	0.01	0.12	0.42	0.18

Installation

Please refer to INSTALL.md for the installation of UniLION codebase.

Data Processing

First, download the nuscenes dataset

├── can_bus
├── maps
├── occ3d
├── samples
├── sweeps 
└── v1.0-trainval

The dataset of occ3d can be downloaded from: https://tsinghua-mars-lab.github.io/Occ3D/

Then, we run the following script:

export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH
export PYTHONPATH="$(dirname $0)/..":$PYTHONPATH
python tools/data_converter/nuscenes_converter.py nuscenes \
    --root-path ./data/nuscenes \
    --canbus ./data/nuscenes \
    --out-dir ./data/nuscenes \
    --extra-tag nuscenes \
    --db-save-path ./data/nuscenes/ \
    --version v1.0

For planning anchor generation, please run:

python kmeans_planning.py

Dwonload image pretrained model

We adopt the same image pretrained model with bevfusion. You could download this file in swint_nuimg_pretrained.pth

After completing all the steps above, the final data structure should be organized as follows:

├── can_bus
├── kmeans_motion_6.npy
├── kmeans_planning_4096.npy
├── kmeans_planning_6.npy
├── maps
├── nuscenes_dbinfos_train.pkl
├── nuscenes_gt_database
├── nuscenes_infos_train.pkl
├── nuscenes_infos_val.pkl
├── occ3d
├── samples
├── sweeps 
└── v1.0-trainval

Getting Started

Train UniLION on nuScenes

# First stage

tools/dist_train.sh projects/unilion_swin_384_det_map.py 8

# Second stage

tools/dist_train.sh projects/unilion_swin_384_seq_perception.py 8

# Third stage

tools/dist_train.sh projects/unilion_swin_384_seq_e2e.py 8

Note: for better performance, you can additionly train 3D object detection for providing the pretrained model in the first stage.

tools/dist_train.sh projects/unilion_swin_384_det.py 8

Test UniLION on nuScenes

You can freely select the tasks to be evaluated in the config.

tools/dist_test.sh projects/<CONFIGS> 8 <CKPT> --eval mAP

Besides, you can use our released model (UniLION model) to evaluate all results:

tools/dist_test.sh projects/projects/configs/unilion_swin_384_seq_e2e.py 8 <CKPT> --eval mAP

TODO

• Release the paper.
• Release the code of UniLION.
• Release checkpoints of UniLION.

Citation

@article{liu2024lion,
  title={LION: Linear Group RNN for 3D Object Detection in Point Clouds},
  author={Zhe Liu, Jinghua Hou, Xingyu Wang, Xiaoqing Ye, Jingdong Wang, Hengshuang Zhao, Xiang Bai},
  journal={Advances in Neural Information Processing Systems},
  year={2024}
}

@article{liu2025unilion,
  title={UniLION:Towards Unified Autonomous Driving Model with Linear Group RNNs},
  author={Zhe Liu, Jinghua Hou, Xiaoqing Ye, Jingdong Wang, Hengshuang Zhao, Xiang Bai},
  journal={arXiv preprint arXiv:2511.01768},
  year={2025}
}

Acknowledgements

We thank these great works and open-source repositories: LION, MMDectection, SparseDrive, Mamba, RWKV, Vision-RWKV, and flash-linear-attention.