NeurIPS 2025 – Conference trip for the PokéAgent Challenge

7 December 2025, by Arian Urdu

Photo: Arian Urdu

As part of Team Hamburg Pokérunners (Benedikt Schink, Matin Urdu, Arian Urdu), we achieved 2^nd place in the second track of the international PokéAgent Challenge (pokeagent.github.io). As a result, one team member was invited to deliver a Lightning Talk at a workshop during NeurIPS 2025 (Conference on Neural Information Processing Systems). I, Arian Urdu, took on this role and traveled to San Diego for the conference, with a generous travel grant by the Claussen-Simon-Stiftung.

The challenge involved training a neural network to complete the game Pokémon Emerald as quickly as possible, up to the first gym badge. Our approach relied on Reinforcement Learning (RL), using the Proximal Policy Optimization (PPO) algorithm. With this method, our agent achieved the first gym badge after 01:14:43 (hh:mm:ss).

The conference took place from December 2–7 in San Diego and featured an extensive program of presentations, workshops, and expert discussions on the latest advancements in AI research.

Highlights included a keynote by Rich Sutton on the Oak Architecture, where he outlined a potential long-term architecture for AGI and explained why he believes current approaches may not lead to general intelligence, and poster sessions, offering in-depth discussions with authors about their work.

A particularly valuable experience was the personal exchange with fellow researchers and the organizers of the PokéAgent Challenge, especially Seth Karten and Jake Grigsby, who dedicated immense effort to organizing the event.

For me, the trip was especially impactful: I met many inspiring and open-minded researchers, engaged in fascinating conversations, and gained clarity on my future research direction. It also highlighted how advanced the state-of-the-art in other AI fields has become.

I extend my special thanks to base.camp, HCDS, and the Claussen-Simon-Stiftung for their support, which made this participation possible.