This work presents a study for building a Deep Vision pipeline suitable for the Robocup Standard Platform League, a humanoid robot soccer tournament. Specifically, we focus on end-to-end trainable object detection for effective perception using Aldebaran NAO v6 robots. The implementation of such a detector poses two major challenges, those of speed, and resource-effectiveness with respect to memory and computational power. We benchmark architectures using the YOLO and SSD detection paradigms, and identify variants that are able to achieve good detection performance for ball detection, while being able to perform rapid inference. To add to the training data for these networks, we also create a dataset from logs collected by the UT Austin Villa team during previous competitions, and set up an annotation pipeline for training. We utilize the above results and training pipeline to realize a practical, multi-class object detector that enables the robot's vision system to run at 35 Hz while maintaining good detection performance.

PDF

In Proceedings of the RoboCup Symposium, 2022, Bangkok, Thailand, July 2022.

@inproceedings{RoboCup2022-nskiran, title={Towards a Real-Time, Low-Resource, End-to-end Object Detection Pipeline for Robot Soccer}, author={Sai Kiran Narayanaswami and Mauricio Tec and Ishan Durugkar and Siddharth Desai and Bharath Masetty and Sanmit Narvekar and Peter Stone}, booktitle={Proceedings of the RoboCup Symposium, 2022}, month={July}, address={Bangkok, Thailand}, url="http://nn.cs.utexas.edu/?RoboCup2022-nskiran", year={2022} }