Can You Train AI to Stay Safe? Meet us at the Great Exhibition Road Festival

The Formal Methods in AI lab will be joining the Great Exhibition Road Festival 2026 on Saturday 6 and Sunday 7 June, 12:00-18:00, for a weekend of hands-on science and arts activities across Imperial's South Kensington campus.

Our exhibit, Can You Train AI to Stay Safe?, invites visitors to try a live Reinforcement Learning training challenge. Choose Pac-Man or MetaDrive, tune the trade-off between reward and safety cost through a multiple choice quiz, watch the agent train, then compare your final score on the leaderboard. The aim is simple to play with and serious underneath: can an AI system perform well while still respecting hard safety constraints?

The demo introduces Constrained Reinforcement Learning through choices visitors can see and change: e.g. future focus (which corresponds to the discount factor $\gamma$ in RL), costs of safety mistakes, traffic patience, learning speed, and policies that range from reward-chasing to safety-first. These ideas connect directly to real questions in AI safety, in particular with respect to constrained optimisation.

Members of the lab will be at our booth, including: Omar Adalat, Alex Goodall, and Terrence Fernandes. We'd be happy to talk about the demo, our research more broadly, and how formal methods can help make AI systems safer and more trustworthy!

Update: thanks to the public for their participation! We had 218 people try the activity out, with 118 people choosing to play MetaDrive, and 100 playing PacMan. In the next event, we'll try to have even more laptops available to get even more people trying out the activity 🙂