Our project is a compound AI model for playing the online video game slowroads.io. We experimented with various data representations for the game state and found that using the road’s left and right coordinates along with the car’s position and velocity gave the best results. A simple CNN with edge-detection kernels extracts the road boundaries, while another model detects the car’s velocity from the game’s display. This data is then fed into an LLM to decide actions like moving forward or turning. Beyond performance, we compared multiple data formats to identify the most effective structure for agentic reasoning and introduced agent-to-agent context sharing, allowing models trained on separate sessions to exchange learned patterns—leading to more cohesive, adaptive, and human-like driving behavior.
Team
Placement
Participant
