Jonathan Moon

January 2022 - May 2022

For our final project we worked with a WidowX robot from Interbotix and trained it to flip pancakes! Our setup included 3 realsense cameras, AR tags mounted on pancakes, a custom 3D printed frying pan, and a cage surrounding the robot. By collecting rosbags of joint states and velocities during demo flips we were able to train a model that took in the joint/velocity states of the robot and the pose of the pancake (poses were obtained with AR tags) and outputted velocity commands for each joint. Here's a video of our project: Pancake Flipping Robot

August 2021 - December 2021

For the final projects for a functional programming and an AI course, we set out to create an AI to play the viral board game Machi Koro. For the functional programming component, the video game engine was coded in OCaml. The game engine was built to be modular and extensible so that we could easily add new features and game mechanics that come with the game's many expansion packs. The game engine was designed to support multiple players, save states, and customizable game rules. For the AI component, we implemented a modified Monte Carlo Tree Search (MCTS) algorithm to play the game. The AI was trained using self-play and was able achieve superhuman performance, beat human players nearly everytime. The AI was also able to play with different game rules and expansion packs. We used our own heuristics about the game to guide the AI's value function, and simulated multiple AI agents with different configurations to find the best performing agent. While exploring every possible option is a critical part of Monte Carlo Tree Search, it was decided that some situations would never occur in a real game. The anomalous situations arose when a player would own hundreds of coins and have only a couple remaining landmarks to purchase. We would eliminate these unrealistic playouts to make the AI much more efficient.