Executive Summary

We demonstrated our AI agent May's capabilities in Pokémon Emerald, surpassing previous AI performance benchmarks by beating the first three gyms in just 5 hours and 15 minutes. This achievement showcases our AI's ability to handle complex, open-ended gameplay environments while maintaining real-time interaction with users.

About Pokémon Emerald

Pokémon Emerald is a complex role-playing game that serves as a benchmark for testing AI capabilities. The game requires players to navigate an open world, manage inventory, build and train teams, and employ strategic thinking in battles. These elements make it an ideal testing ground for evaluating AI systems, particularly in reinforcement learning applications.

The Challenge

Prior to our attempt, the established record for AI performance in Pokémon Emerald was reaching the second gym. This benchmark had stood as a significant challenge due to the game's inherent complexity. The game demands mastery of open-world exploration, team building, resource management, and strategic battle decisions. For context, human players typically require 2-3 hours just to reach the second gym.

Results

We conducted three livestreamed sessions to demonstrate May's capabilities, completing Gym 1 in 1 hour 45 minutes, Gym 2 in 2 hours 25 minutes, and Gym 3 in 1 hour, for a total completion time of 5 hours 15 minutes. A unique feature of our approach was implementing real-time viewer interaction, allowing stream viewers to influence May's behavior through chat commands.

Here we want to show some key Highlights:

Prepping for expected Battle (long-term planning)

May repeatedly showed smart long-term planning by buying Pokéballs and healing items when anticipating big fights. Check out this screenshot of May stocking up (or find the moment in the VOD here):

Swapping Pokémon and Move Usage

In this highlight we show that May is capable of smart and interpretable reasoning - you can literally read her mind! In this highlight, May switched her active Pokémon from 'Growly' to 'Marshstomp' since Marshstomp has a type advantage, and a move much more effective against the current opponent. (Here is the moment in the stream)

Listening on user Instructions

In the second livestream, a viewer told May there was an NPC giving away an lv. 100 Alakazam. Even though this was a complete lie, May bought it and went off to talk to every NPC in town, in hopes of finding them! The screenshot below shows how she adapted her reasoning, or you can find the moment in the VOD here.

Outcome

• World Record Achievement: May successfully completed all three gyms in 5 hours and 15 minutes, surpassing the previous AI record of two gyms.
• Real-time Strategy Demonstration: The livestream showcased May's decision-making process, giving users direct insight into the AI's reasoning and adaptability.
• Interactive AI Development: May successfully handled user commands while maintaining progress toward game objectives, proving the viability of interactive AI agents.