RoboJEPA reportedly shows scaling laws for robot world models

A post says RoboJEPA’s authors fit a scaling law to models ranging from 22 million to 2 billion parameters, then accurately predicted results for 4B and 8B models. It reports that the 8B model grasped objects 67% of the time on a real Franka robot without task fine-tuning, versus 5% for π0.5. The post notes that the comparison used a goal image versus text; π0.5 did better at pick-and-place, 53% versus 27%.