Developing AI agents that work together with humans to solve problems could help us to make progress on some of the most complex problems we face, but it requires experiments that recreate some of the messiness and open-endedness of real-world scenarios. Collaborative AI aims to address this challenge by developing smart machines that can learn to co‑operate with people and other AI agents. But understanding what drives individual and collective behaviours when multiple participants interact in complex environments is not easy. It requires AI to make sense of incomplete or noisy information and make sequential decisions. This may be further complicated by individuals’ choices about whether to compete or co‑operate with others.

Project Malmo is an open-source platform for asking research questions about multi-agent interactions. This platform is built on top of the Minecraft gaming environment, which mimics the limitless and spontaneous qualities of the real world, allowing researchers to investigate scenarios that are relevant to solving real-world problems. This aim is the guiding principle behind the Project Malmo Challenges.

In 2017, Microsoft Research ran their first competition on the platform: the Malmo Collaborative AI Challenge. They invited teams worldwide to develop and submit the code for artificial agents to complete a game called Catch the Pig. To win this game, AI agents had to work collaboratively with other artificial and human players. Competing teams tried a wide range of AI methods and learning paradigms to make agents with the best collaborative strategies.

There are many different approaches to multi-agent learning and interactions, and it is difficult to predict which ones might be optimal in a given situation. To ensure that researchers make timely progress, it is sometimes necessary to incentivise many different groups to all work on the same problem. During the two-month competition window, the Challenge received entries from 80 teams across 26 countries, identifying many new approaches to collaborative AI. Project Malmo has continued to be used as a testing environment, with two new Challenges launched in 2019 that are further helping the research community to better understand the interactions of multiple agents within systems. In this example, we see collective intelligence being used to stimulate research into AI methods, which can directly help us to optimise hybrid human–machine systems that collaborate towards a shared goal.