AutoMine Solution for AV2 2026 Scenario Mining Challenge

🏆 Competition   |   📄 Technical Report

Songliang Cao^{1,2 *}, Jiele Zhao^{1 *}, Yuru Wang¹, Hao Li¹, Daqi Liu¹, Zehan Zhang^{1 †}
Fangzhen Li^{1 †}, Yu Wang, Yue Zhang, Bing Wang¹, Guang Chen¹, Hao Lu², Hangjun Ye¹

¹Xiaomi EV   ²Huazhong University of Science and Technology

^*This work was done during internship at Xiaomi.   ^†Project leader.

Table of Contents

• 1. Installation
• 2. Data Preparation
  • 2.1 Argoverse2 Sensor Dataset
  • 2.2 Argoverse2 Scenario Mining Dataset
  • 2.3 Tracking Predictions
• 3. Running
  • 3.1 Track Refinement
  • 3.2 Scenario Mining
• Acknowledgements

1. Installation

Using Conda is recommended for environment management

conda create -n refAV python=3.10
conda activate refAV

All of the required libaries and packages can be installed with

pip install -r requirements.txt
export PYTHONPATH=.

2. Data Preparation

2.1 Argoverse2 Sensor Dataset

For more information check out Argoverse User Guide.

conda install s5cmd -c conda-forge

export DATASET_NAME="sensor"  # sensor, lidar, motion_forecasting or tbv.
export TARGET_DIR="$HOME/data/datasets"  # Target directory on your machine.

s5cmd --no-sign-request cp "s3://argoverse/datasets/av2/$DATASET_NAME/*" $TARGET_DIR

2.2 Argoverse2 Scenario Mining Dataset

The RefAV scenario mining dataset can be downloaded from Huggingface using:

pip install huggingface_hub
hf auth login

export TARGET_DIR="$(pwd)/scenario_mining_downloads"
hf download CainanD/RefAV --repo-type dataset --local-dir=$TARGET_DIR

Alternatively, you may download the scenario-mining add on from Argoverse without having to sign-in to Huggingface with:

export TARGET_DIR="$(pwd)/scenario_mining_downloads"
s5cmd --no-sign-request cp "s3://argoverse/tasks/scenario_mining/*" $TARGET_DIR

2.3 Tracking Predictions

Our method is based on 3D tracking results. You can either generate tracking results using a SOTA method, or download existing results (our choice):

• Option 1 — Generate your own tracking results. See the LT3D repository for information on training a baseline detector and tracker on the Argoverse 2 dataset.
• Option 2 — Download existing tracking results (recommended). We use the tracking outputs from previous winning Argoverse submissions:
  • Test set: Google Drive
  • Val set: Hugging Face or Google Drive

3. Running

Before running the code, please configure the relevant paths in refAV/path.py.

3.1 Track Refinement

The provided 3D tracking results contain a certain amount of noise, so we refine them before scenario mining.

Note: The validation-set tracking results contain a significant number of over-detections, which slow down the pipeline and degrade final performance. We use a VLM for grounding to filter out most of these over-detections.

# use Qwen3.5 for grounding
git clone https://github.com/QwenLM/Qwen3-VL.git grounding/Qwen3-VL
python grounding/batch_grounding_av2.py

# correct the tracking results using the grounded results
python grounding/correct_tracking.py

For both the test and val tracking results, we additionally apply a unified track-refinement step to mitigate trajectory noise (ID switches, overlapping boxes, etc.):

python grounding/run_tracker.py 
        --val_dir /path/to/val \
        --output_dir /path/to/output \
        --mode stitch

For convenience, we also provide the refined tracking results so you can skip this step: TODO.

3.2 Scenario Mining

Based on the refined tracking results, we perform scenario mining and generate the scenario description files.

A. Description augmentation

python desc_aug.py 
  --batch_file tools/log_prompt_pairs_test_unique.json \
  --output_json output/desc_aug/desc_aug_test.json \
  --model claude-opus-4-6 

B. Generate base code with double check

python double_check_desc_aug.py 
  --aug_file output/desc_aug/desc_aug_test.json  
  --output_dir output/llm_code_predictions/RefAV/double_check_test 
  --include_original
  --workers 12 

This produces generated code for each description-augmentation variant, written to double_check_test/, double_check_test_aug1/, and double_check_test_aug2/ respectively.

C. Self-iterative refinement

For each augmentation variant, we run a self-iterative feedback loop that re-prompts the coder model with execution feedback to refine the generated code. Run the command below once per variant by changing the --base argument to double_check_test, double_check_test_aug1, and double_check_test_aug2 in turn.

bash iterative_feedback.sh \
  --base double_check_test \  # repeat with double_check_test_aug1 / double_check_test_aug2
  --rounds 5 \
  --split test \
  --tracker Le3DE2D_tracking_tz_adjustment_refine \
  --atomic_func_name atomic_functions_0529 \
  --batch_file tools/log_prompt_pairs_test_unique.json \
  --log_prompts /data/Scenario_Mining_Challenge/av2_sm_downloads/log_prompt_pairs_test.json \
  --coder claude-opus-4-6

D. Ensemble prediction

We ensemble the refined predictions from all three augmentation variants to produce the final submission.

python pkl_ensemble.py

Acknowledgements

Our work is built upon the foundational framework provided by the RefAV repository. We extend our sincere gratitude to the authors for making their code public.