Home Original page

GitHub - ucla-mobility/AutoVLA: [NeurIPS 2025] AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning

[NeurIPS 2025] This is the official implementation of the paper:

AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning

Zewei Zhou*, Tianhui Cai*, Seth Z. Zhao, Yun Zhang, Zhiyu Huang†, Bolei Zhou, Jiaqi Ma

University of California, Los Angeles | * Equal contribution, † Project leader

teaser

  • 🚗 AutoVLA integrates chain-of-thought (CoT) reasoning and physical action tokenization to directly generate planning trajectories through a unified autoregressive generative process, dynamically switching thinking modes.
  • ⚙️ Supervised fine-tuning (SFT) is employed to enable the model with dual thinking modes: fast thinking (trajectory-only) and slow thinking (enhanced with CoT reasoning).
  • 🪜 Reinforcement fine-tuning (RFT) based on Group Relative Policy Optimization (GRPO) is adopted to enhance planning performance and runtime efficiency, reducing unnecessary reasoning in straightforward scenarios.
  • 🔥 Extensive experiments across real-world and simulated datasets and benchmarks, including nuPlan, nuScenes, Waymo, and CARLA, demonstrate its competitive performance in both open-loop and closed-loop settings.

News

  • 2026/02: AutoVLA codebase is now released.
  • 2025/09: AutoVLA is accepted by NeurIPS 2025 👏👏.
  • 2025/06: AutoVLA paper release.
  • 2025/05: In the Waymo Vision-based End-to-end Driving Challenge, AutoVLA ranks highly in both RFS Overall and achieves the top RFS Spotlight score, which focuses on the most challenging scenarios.

Release Plan

  • 2025/06: ✅ AutoVLA paper.
  • 2026/02: ✅ AutoVLA annotation and training code.
  • 2026/03: AutoVLA checkpoints.
  • TBD : Reasoning data (Pending approval from the data provider).

Devkit Setup

1. Dataset Downloading

nuPlan Dataset

You can refer to here to prepare the nuPlan dataset. Be careful with the dataset structure.

bash navsim/download/download_maps.sh
bash navsim/download/download_trainval.sh
bash navsim/download/download_test.sh

Waymo E2E Dataset

The Waymo end-to-end driving dataset can be downloaded at here.

nuScenes Dataset

The nuScenes dataset can be downloaded from the official website: https://www.nuscenes.org/. You will need to register and download the v1.0-trainval split.

2. Conda Environment Setup

You can perform the following command to create a conda environment and install the required dependencies.

conda env create -f environment.yml
conda activate autovla
pip install -e . --no-warn-conflicts
bash install.sh

3. Navsim Setup

We have included the navsim code in this repo, and you can go to the navsim folder to install it. You can also refer to here to set up the navsim devkit, but please ensure version compatibility for the dependencies.

cd navsim
pip install -e . --no-warn-conflicts

Remember to set the navsim required environment variables:

export NUPLAN_MAP_VERSION="nuplan-maps-v1.0"
export NUPLAN_MAPS_ROOT="$HOME/navsim_workspace/dataset/maps"
export NAVSIM_EXP_ROOT="$HOME/navsim_workspace/exp"
export NAVSIM_DEVKIT_ROOT="$HOME/navsim_workspace/navsim"
export OPENSCENE_DATA_ROOT="$HOME/navsim_workspace/dataset"

4. Pretrained Model Downloading

We use the Qwen2.5-VL model series as the pretrained VLM in the VLA model and CoT annotation model. You can run the command to download the pretrained model.

bash scripts/download_qwen.sh

Specifically, we use the 72B model in CoT annotation, and you can choose Qwen2.5-VL-72B-Instruct or Qwen2.5-VL-72B-Instruct-AWQ based on your device. We use the Qwen2.5-VL-3B-Instruct in the AutoVLA model.

Getting Started

1. Dataset Preprocessing

nuPlan Dataset

You can perform the command to preprocess the nuPlan dataset. Please first revise your path and data split (refer to here) in the config. The INCLUDE_COT setting in the bash determines whether to launch the CoT reasoning annotation.

bash scripts/run_nuplan_preprocessing.sh

Waymo E2E Dataset

To organize the image data and support random access, we first cache the image data in the same format as the other dataset we used.

bash scripts/run_waymo_e2e_image_extraction.sh

You can perform the following command to preprocess the Waymo E2E dataset. Please also first revise your path and data split in the config and set the INCLUDE_COT.

bash scripts/run_waymo_e2e_preprocessing.sh

You can use waymo_e2e_traj_project_visualization.py and waymo_e2e_visualization.py in the tools/visualization folder to visualize the Waymo data after preprocessing.

nuScenes Dataset

You can download the DriveLM nuScenes annotations (v1_1_train_nus.json) from https://github.com/OpenDriveLab/DriveLM/tree/main/challenge.

Note: nuScenes preprocessing requires nuscenes-devkit, which might have dependency conflicts with the main environment. We recommend using a separate conda environment:

# Create a separate environment for nuScenes preprocessing
conda env create -f environment_nusc_preprocess.yml
conda activate nusc_preprocess

# Run preprocessing
bash scripts/run_nuscenes_preprocessing.sh \
    --nuscenes_path /path/to/nuscenes \
    --output_dir /path/to/output \
    --drivelm_path /path/to/drivelm/v1_1_train_nus.json

# Switch back to the main environment when done
conda activate autovla

2. Action Codebook Creation

The action codebook discretizes continuous vehicle trajectories into a finite vocabulary for autoregressive prediction. To create the codebook from your preprocessed data:

python tools/action_token/action_token_cluster.py \
    --data_path /path/to/preprocessed/nuplan/data \
    --output codebook_cache/agent_vocab.pkl \
    --num_cluster 2048

This will generate a vocabulary file that maps trajectory segments to discrete tokens.

3. Supervised Fine-tuning (SFT)

First revise the dataset path and SFT parameters in the config file in config/training. You can customize:

  • data.train.json_dataset_path: Dataset paths for training (supports multiple datasets as a list)
  • data.train.sensor_data_path: Corresponding sensor data paths
  • training.train_sample_size: Set to a number to train on a random subset, or null for the full dataset
  • model.use_cot: Enable/disable chain-of-thought reasoning in training data

Then, launch the SFT training:

python tools/run_sft.py --config training/qwen2.5-vl-3B-mix-sft

4. Reinforcement Fine-tuning (RFT)

You can revise your dataset path and GRPO parameters in the config file in config/training. Then, execute the following command to run reinforcement finetuning.

5. Evaluation

nuPlan Evaluation (Navsim)

We leverage Navsim and its Predictive Driver Model Score (PDMS) to test and evaluate our model on nuPlan. You need to set up the dataset path and split in the evaluation bash, and run the command to launch the testing.

bash navsim/scripts/evaluation/run_autovla_agent_pdm_score_evaluation.sh

nuScenes Evaluation

To evaluate the AutoVLA model on nuScenes validation data, you need to prepare the segmentation data for collision evaluation. You can download the preprocessed segmentation data from this link, which we preprocessed using code from UniAD.

Then run:

python tools/eval/nusc_eval.py \
    --config config/training/qwen2.5-vl-3B-nusc-sft.yaml \
    --checkpoint /path/to/checkpoint.ckpt \
    --seg_data_path /path/to/nusc_eval_seg

Citation

If you find this repository useful for your research, please consider giving us a star 🌟 and citing our paper.

@article{zhou2025autovla,
 title={AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning},
 author={Zhou, Zewei and Cai, Tianhui and Zhao, Seth Z.and Zhang, Yun and Huang, Zhiyu and Zhou, Bolei and Ma, Jiaqi},
 journal={Advances in Neural Information Processing Systems (NeurIPS)},
 year={2025}
}