DeepMind's RoboCat learns and improves across different robot bodies

Google DeepMind announced RoboCat, a robotic agent built on Gato, the lab’s earlier multimodal “generalist” model that could process language, images and actions across simulated and physical environments. RoboCat could operate several different types of robot arm, including a three-fingered gripper, and learn new manipulation tasks from a small number of human demonstrations.

The system trained through a self-improvement loop rather than a single fixed dataset. DeepMind collected 100 to 1,000 human-controlled demonstrations of a task, fine-tuned a specialised version of RoboCat on them, then had that specialised agent attempt the task on its own around 10,000 times to generate new training data. The resulting mix of human and self-generated data was folded back into the main training set to produce an updated version of RoboCat, and the cycle repeated — DeepMind described this as “a virtuous cycle of training” in which each new task learned made the next one easier to pick up. The lab reported that RoboCat could adapt to a new, previously unseen robot arm within a few hours of practice, and that iterating the self-improvement cycle raised its success rate on unseen tasks from 36% to 74%; in one manipulation benchmark, picking up gears, it reached an 86% success rate after 1,000 demonstrations.

RoboCat sat within DeepMind’s broader effort to build robot-control systems that generalised across hardware rather than being retrained from scratch for each new arm or gripper — a persistent bottleneck in robotics research. The same self-generated-data approach anticipated methods DeepMind would later fold into its RT-2 vision-language-action model the following month.