Paper
Publication
A Generalist Agent
Abstract

Inspired by progress in large-scale language modelling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy. The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens. In this report we describe the model and the data, and document the current capabilities of Gato.

Authors' notes

Inspired by progress in large-scale language modelling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy. The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens.

During the training phase of Gato, data from different tasks and modalities are serialised into a flat sequence of tokens, batched, and processed by a transformer neural network similar to a large language model. The loss is masked so that Gato only predicts action and text targets.

When deploying Gato, a prompt, such as a demonstration, is tokenised, forming the initial sequence. Next, the environment yields the first observation, which is also tokenised and appended to the sequence. Gato samples the action vector autoregressively, one token at a time.

Once all tokens comprising the action vector have been sampled (determined by the action specification of the environment), the action is decoded and sent to the environment which steps and yields a new observation. Then the procedure repeats. The model always sees all previous observations and actions within its context window of 1024 tokens.

Gato is trained on a large number of datasets comprising agent experience in both simulated and real-world environments, in addition to a variety of natural language and image datasets. The number of tasks, where the performance of the pretrained Gato model is above a percentage of expert score, grouped by domain, is shown here.

The following images also show how the pre-trained Gato model with the same weights can do image captioning, engage in an interactive dialogue, and control a robot arm, among many other tasks.