NaoQNN: Quantized Neural Networks for NAO V6 robots

This repository contains code for the paper "Quantized Neural Networks for Ball Detection on the NAO Robot: An Optimized Implementation", presented at the RoboCup 2024 Symposium.

Paper

Thielke, F. (2025). Quantized Neural Networks for Ball Detection on the NAO Robot: An Optimized Implementation. In: Barros, E., Hanna, J.P., Okada, H., Torta, E. (eds) RoboCup 2024: Robot World Cup XXVII. RoboCup 2024. Lecture Notes in Computer Science(), vol 15570. Springer, Cham. https://doi.org/10.1007/978-3-031-85859-8_9

How to reproduce results from the paper

1. Open ball_model_experiments and either start the devcontainer or install NaoQNN and requirements.txt into your Python environment.
2. Download b-alls-2019.hdf5.
3. Run python prepare_dataset.py to prepare the TensorFlow datasets.
4. Train the models by executing python train_ball_model.py 0 to python train_ball_model.py 7. Since the models are very small, depending on your hardware these 8 processes can easily run in parallel on the same machine.
5. Run python inference_test.py to perform predictions on the test set.
6. Run test_models/export_models.py to convert the quantized models to asm and the float models to H5 files.
7. Compile test_models/measure_speed.cpp and link it with CompiledNN. Run the resulting executable on a NAO V6 robot to get inference timings.
8. Check that the resulting inference times match those in MEASURED_INFERENCE_TIME in calculate_metrics.py and correct them if necessary.
9. Run python calculate_metrics.py to generate plots and a LaTeX table with the results.
10. Run python plot_ball_detections.py to recreate Fig. 4 from the paper. Note that by default the plotted cases are chosen randomly.