Jack of All Trades, Master of Some, a Multi-Purpose Transformer Agent

Hugging Face · 2 years ago · 47

Researchers released JAT, a single transformer model trained to perform multiple tasks across video games, robotics, and language understanding by learning from expert agents. The model achieved 65.8% average performance compared to expert agents across four domains, with particularly strong results on BabyAI (99.0% of expert score) and MuJoCo (84.8% of expert score). The project released a dataset containing hundreds of thousands of expert trajectories and expert policies across 157 training tasks, enabling future development of generalist agents.