|[📖 Paper] [🤗 AutoDrive-R2-7B] [🤗 AutoDrive-R2-7B-COT-SFT] [📊 All Data]

AutoDrive-R2 is a specialized Vision-Language-Action (VLA) model designed for autonomous driving trajectory prediction, which elicits reasoning and self-reflection capacities through rule-based Reinforcement Learning (RL).

Given a front-view camera image and historical vehicle status (position, velocity, acceleration, steering angle), AutoDrive-R2 predicts future waypoints at 0.5s intervals for the next 3 seconds.

Input:

• Front-view camera image
• Historical vehicle status (last 2.0-3.0 seconds at 0.5s intervals)

Output:

• 6 future waypoints: [x(t+0.5s), x(t+1.0s), ..., x(t+3.0s)]
• Each waypoint: [x_coordinate, y_coordinate] in meters

Reasoning Process:

1. Visual Analysis: Analyze road conditions, obstacles, and traffic signals
2. Motion Modeling: Apply kinematic equations for trajectory prediction
3. Logical Deductions: Safety checks and path planning
4. Self-Reflection: Validate predicted trajectory feasibility

Pipeline of our method. We adopt a two-stage training process. The first stage introduces an innovative CoT dataset named nuScenesR²-6K for SFT. The nuScenesR²-6K adopts a four-step logical chain with self-reflection to generate valuable chain-of-thought data. The second stage proposes an novel physics-grounded reward framework for RL optimization, which incorporates spatial alignment, vehicle dynamic, and temporal smoothness for reliable trajectory planning.

• Qwen2.5-VL Base Model: Leverages state-of-the-art vision-language capabilities
• Chain-of-Thought Reasoning: Explicit reasoning steps for interpretable predictions
• Rule-based RL Training: GRPO with physics-grounded rewards
• Multi-dataset Support: Trained and evaluated on nuScenes and Waymo

Each sample contains historical vehicle status at 0.5s intervals:

• Position: [x, y] coordinates (x=forward, y=leftward)
• Velocity: v in m/s
• Acceleration: a_x, a_y in m/s²
• Steering angle: θ (positive=left turn, negative=right turn)

Download from nuScenes Official Website:

nuscenes/
├── samples/
│   ├── CAM_BACK/
│   ├── CAM_BACK_LEFT/
│   ├── CAM_BACK_RIGHT/
│   ├── CAM_FRONT/
│   ├── CAM_FRONT_LEFT/
│   └── CAM_FRONT_RIGHT/
└── sweeps/
    ├── CAM_BACK/
    ├── CAM_BACK_LEFT/
    ├── CAM_BACK_RIGHT/
    ├── CAM_FRONT/
    ├── CAM_FRONT_LEFT/
    └── CAM_FRONT_RIGHT/

Download from Baidu Netdisk: [Link Coming Soon]

# Clone the repository
git clone https://github.com/your-repo/AutoDrive-R2
cd AutoDrive-R2

# Create conda environment
conda create -n autodrive-r2 python=3.11
conda activate autodrive-r2

# Install dependencies
pip install -r requirements.txt

# Install flash-attention
pip install flash-attn --no-build-isolation

Generate Chain-of-Thought reasoning for SFT training:

python AScripts/cot.py

Perform SFT training on CoT-annotated data:

bash src/scripts/run_sft_video_7B_6k.sh

Fine-tune with GRPO using physics-grounded rewards:

bash src/scripts/run_grpo_video_7B_6k.sh

python AScripts/eval_nuscene.py

python AScripts/eval_waymo.py

If you find our work helpful for your research, please consider citing:

@inproceedings{
    yuan2026autodriver,
    title={AutoDrive-R{\texttwosuperior}: Incentivizing Reasoning and Self-Reflection Capacity for {VLA} Model in Autonomous Driving},
    author={Zhenlong Yuan and Chengxuan Qian and Jing Tang and Rui Chen and Zijian Song and Lei Sun and Xiangxiang Chu and Yujun Cai and Dapeng Zhang and Shuo Li},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=KVWaCzJrrq}
}

We sincerely appreciate the contributions of the open-source community:

• Video-R1
• lmm-r1
• DeepSeek-R1
• Qwen2.5-VL
• nuScenes
• Waymo Open Dataset