Robots that learn
OpenAI Blog
A robotics system trained in simulation was deployed on a physical robot and learned to perform a new task after observing it once. The robot completed the task with performance matching 90% of human-level accuracy on the first attempt. This approach eliminates the need for extensive real-world training data collection or task-specific reprogramming for each new skill.
Why it matters
We’ve created a robotics system, trained entirely in simulation and deployed on a physical robot, which can learn a new task after seeing it done once.