world labs
Long-horizon tasks in interactive worlds.
World Labs is a research lab advancing agents through interactive worlds.
We are building a continuing supply of original, high-quality training data alongside reinforcement-learning environments in real game engines.
Our work connects two offerings for AI labs:
- Original datasets for pretraining and continued pretraining, centred on technical correctness, novelty, and depth.
- Game-engine gyms for post-training: executable problem sets with resettable projects, engine tools, verifiers, and reward signals.
Game worlds connect code, spatial structure, and runtime behavior. Our direction spans environments such as Roblox Studio, Unreal Engine, and Minecraft, with tasks that require sustained planning, debugging, and recovery across interacting systems.
Our research loop uses controlled training runs and held-out gym projects to measure what new data teaches a model. We evaluate completion, regressions, and recovery, then use the results to inform the next dataset and environment release.
Game worlds for agent training
We’re exploring partnerships with game developers to adapt existing games into training and evaluation environments for AI agents. We’re interested in worlds that demand sustained planning, interaction, and recovery. Pilot work would focus on agent interfaces, repeatable starting states, and verifiable outcomes, with private deployment requirements shaped alongside research partners.
For data partnerships and gym pilots, get in touch at dhruv@tryworldlabs.com.