Hi everyone!

Remember how we showed off our AI completing the Intro Tunnel of Hogwarts Legacy in December? Well, that was cool, but the storyline there is pretty linear. We wanted to really push our AI's limits, so we took on a tougher challenge: Pokémon Emerald.

Pokémon Emerald and the past world record

For years, research teams have been using Pokémon Emerald to test AI systems, especially for reinforcement learning (RL). It's a great benchmark because it tests all sorts of AI capabilities. Until now, the best RL-based AI bots could only reach the second gym - that was the world record. To get this far, the player needs to explore the map, build a team of Pokémon, and manage their inventory, among other challenges. For context, most humans need about 2-3 hours to get that far.

Our World Record Stream

To show off what our AI agent May can do in more open-ended games, we decided to livestream everything. We did three streams, one for each gym:

Gym 1: 1h45min of gameplay
Gym 2: 2h25min of gamplay
Gym 3: 1h of gamplay

That's right, we broke the record! 🎉 May reached and beat the third gym in just 5h15min total. That's actually faster than I was when I first played Pokémon!

Interactability

Here's a cool underappreciated feature: viewers could change May's behavior during the stream by typing commands in the chat (like "!walk a circle"). May would then try to follow these instructions. We've included some of these moments in our highlights.

Highlights

1. The Great Alakazam Hunt

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.

2. 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):

3. Swapping Pokémon and Move Usage

The last highlight we want to showcase is 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)

What This Means

By beating the third gym, May has become the world's best AI for Pokémon as of writing. The cool part is we did this live, so nothing was cherry-picked. We're excited about what this means for AI in gaming and beyond - from game testing to player simulation.

Think your project could benefit from AI like May? Let's talk!

Reach out to us and see how our advanced AI can take your work to the next level.