MT-OPT: Google trains a single robot policy on 800+ tasks and 57,000 hours of real data
In one sentence Google pre-trains a single policy on over 800 real robot tasks and 57,000 hours of real-world data, demonstrating for the first time zero-shot transfer to new tasks through large-scale multi-task offline learning.
Imagine teaching someone a trade from scratch. Usually you'd focus on one task at a time. Google did something very different: they had a robot practice over 800 different tasks simultaneously, accumulating 57,000 hours of real-world experience. Not simulated — real.
The result is called MT-OPT, short for Multi-Task Offline Pre-Training. The idea is as simple as it is powerful: if a robot sees enough variation across different tasks during training, it learns general principles it can then apply to tasks it has never seen before, without any additional training.
This is called zero-shot transfer: the robot does not need to start from scratch for every new task. Like a chef who, having cooked thousands of dishes, can intuitively adapt to a new recipe without having to study it for days.
Before MT-OPT, robots were generally trained on individual tasks. This work shows that scale and data diversity matter in robotics just as they do in language models. It is one of the first concrete pieces of evidence that the foundation model paradigm can work in the physical world too.
Companies
Tools
—
Tags
Sources