This is a hobby research project in which we are using a foosball table as a platform for learning about, developing and implementing a deep learning neural network.

We have built a custom robot to interface with a single foosball rod. We capture webcam video of human players to train our FoosAI brain to extract the current rod position and predict the future rod position. The robot moves the rod accordingly.

The project is in the early stages and we have recently had our first successful model operate the foosball robot. I will be maintaining the code and project details on GitHub.

GitHub Project

The source code for FoosAI is available through GitHub. Contributions are welcome.

Videos and Images

The video demonstrates a single foosball rod connected to a custom-built robot actuator that is being controlled by our deep learning neural network.

Foosball table
Fig. 1 - This is the foosball table at Microsoft, Vancouver.
FoosAI schematic
Fig. 2 - A schematic showing our system setup. We capture webcam video of the foosball table, which is the input to the deep learning AI, which in turn controls the robotic foosball rod.
FoosAI robot
Fig. 3 - Photo of our custom robot that interfaces with the foosball rod.
FoosAI robot-arm connection
Fig. 4 - Close-up of the actuator-foosball rod connection.
FoosAI table extraction
Fig. 5 - The blue boxes indicate the rod locations which are used by the deep learning CNN to extract the rod positions.
FoosAI Training Result Position 1 FoosAI Training Result Position 2
Fig. 6 - This is a figure caption.
FoosAI Future Prediction Position 1 FoosAI Future Prediction Position 2
Fig. 7 - This is a figure caption.
FoosAI controlling the foosball rod
Fig. 8 - The model is evaluated in real-time against live webcam data, which is sent over serial to the robot.
Back to top