Researchers at the University of California, Los Angeles (UCLA) have recently introduced Teamcraft, an open platform designed for learning and evaluating AI algorithms in complex multitasking environments. Inspired by the mechanics of the popular game Minecraft, Teamcraft aims to test the interactions of AI agents within teams. The details of the project can be found in the research paper published on arXiv.
Teamcraft offers a range of training and evaluation tasks, including construction, farming, cleaning territories, and transporting materials. The platform supports tasks using multimodal inputs, creating more intricate interaction scenarios. Each AI agent operates in a first-person perspective, receiving data through an RGB camera to mimic human perception.
The platform features various management strategies, supporting both centralized and decentralized coordination approaches for teams. With over 55,000 unique task variations representing different biomes, materials, and conditions, agents can take on roles with specific duties, adjusting to dynamic environments and making real-time decisions. Agents perform predefined actions similar to those seen in Minecraft gameplay.
Developers emphasize the significance of scaling training data for enhancing agents’ abilities to perform complex tasks and collaborate effectively within teams. This scalability opens up possibilities for creating more advanced AI systems in the future.
Teamcraft is now publicly available on GitHub, allowing scientists to utilize the platform to test their own algorithms. The project envisions Teamcraft evolving into a testbed for human-AI interactions, potentially involving real players in gaming environments.