Definition
Plain language
An AI system that learns by predicting the next state of the world, and picked up robot control just from watching video it never acted on.
As stated in the literature
A BAAI world model trained to predict latent world-state transitions bidirectionally from video, with thin frozen-backbone decoders for text, image, and action; shows transfer to robot control without action labels.
Why it matters: It shows a system can learn useful control just from passive observation, which could sharply reduce the need for costly hands-on training data.
For example, after only watching videos of a task without ever moving anything itself, Orca could go on to control a robot doing that task.
Heard on the show
“The paper is called Orca, from the Beijing Academy of Artificial Intelligence, posted at the end of June.”Episode 213 — A Model Learned to Control a Robot by Watching Video It Never Acted On