Jack of All Trades (JAT) is a multi-purpose transformer agent capable of mastering a wide variety of sequential decision-making tasks, as well as showing rudimentary capabilities in natural language processing and computer vision. Key highlights:
– Release of expert reinforcement learning (RL) agents across various environments, including Atari, BabyAI, Meta-World, and MuJoCo.
– Introduction of the JAT dataset, the first dataset for generalist agent training, containing hundreds of thousands of expert trajectories.
– Description of the JAT agent architecture, a transformer-based model that can handle sequential data and continuous values.
– Experimental results showing that JAT can achieve 65.8% average performance compared to expert agents across the 157 training tasks.
– Surprising finding that adding observation prediction as an auxiliary objective can improve the agent’s learning efficiency.
The authors conclude by highlighting future research directions, such as improving the dataset, leveraging offline RL, and exploring smarter multi-task sampling strategies to further unlock the potential of generalist agents.