RoboticsWorld modelsStory 01
One interaction can teach a robot hidden physics
What changedT2-Z adds an environment latent that is optimized from one or more interactions while the trained robotic world model stays frozen. Reported prediction error fell from 15.20 to 7.00 and action success rose from 33.73 to 72.96 percent; some settings adapted from one trajectory.
The useful part
Why it matters
Reported prediction error fell from 15.20 to 7.00 and action success rose from 33.73 to 72.96 percent; some settings adapted from one trajectory.
Keep in mind
Good to know
All results are author-reported.
Evidence