A generalist AI agent for 3D virtual environments

Such research aims to build interactive agents capable of capturing the nuances of situated interactions using multi-agent reinforcement learning, as well as develop generally capable agents able to solve complex game environments by starting from simpler simulations and utilizing demonstrations. AlphaStar, a preliminary version of an agent that achieved the top league in a widely popular esports game, is considered a significant accomplishment for the field. The use of Unity to create more diverse virtual simulations and demonstrate efficient usage of demonstrations are also highlighted.