Cozmo Partial Cube Detector

By Tyler Johnson

Project Description: Partial cube detector which uses a convolutional neural network to classify the left and right images between the classes cube or no-cube. Decisions can then be made based on the prescense of a cube. Cozmo is currently set up to not react to other objects and turn toward cubes.

The results are accurate on the training and testing set, achieving about 93% accuracy on my test set after reaching 99% accuracy during training. The training set was composed of thousands of random cropped images containing cubes and those not containing cubes. The in practice results match the test results fairly well when subjected to similar circumstances as the training. However some common misclassifications happen with similar objects especially around the edge of cozmo's field of view. Further training and experimentation is needed to clean up these false positive cases. Also the models detection accuracy is heavily influenced by the lighting conditions so too dark or bright images present issues and lead to misclassifications.

Detailed design and explanation of the project can be found below, as well as example photos and a video:

The Code:

Summary Powerpoint:

Demo Video: