Tic! Your Tic Tac Toe Playing Robot with Embodied AI

At the 2025 Seeed Emboddied AI Hackathon at Circuit City Mountain View our team attempted to build Tic, a Tic Tac Toe playing robot in under 24 hours.

Initially, we planned to follow the standard le robot setup using the lead follower SO-ARM100, simple USB cameras, and a desktop with an i9 and RTX 3070 for training. The plan was to train a simple ACT model to have something working and prove out the pipeline pivoting to utilizing the foundational robot model called Pi0.

However, we faced several challenges. Our robot didn't come with a camera mount, requiring a new model found and printed to mount the camera. The robot was unstable and crashed if the gripper moved too quickly, a problem we never were able to sort out and eventually just had to be careful not to move too fast while doing data collection. In addition issues with data recording on some computers, resulting in lost time debugging. Finally when attempting to locally train Pi0 it didn't work as expected due to VRAM requirements.

In the end our final approach was to recording 50 expert tic-tac-toe demonstrations evenly distributed over a tic tac toe game, training an ACT model on the data. We used both the 3070 desktop along with a cloud H100 to accelerate training. Out of curiosity we decided to try and have the model learn game rules in addition to the pick and place motion, while a hybrid approach with a classically programmed AI to solve tic tac toe plugged into a grasping ACT model might have been more successful, we wanted to see how well the model would learn with little to no information outside of the expert training.

Overall the weekend was a crazy experience, we all made new friends and got to experience the new world of open source embodied AI. While there is a litany of things we would have done differently in hindsight, we all had a great time and are happy with what we built. See everyone next year for the 2026 Hackathon!

Find more information at our website!