Robotics dataset claimed to cover over one million unique tasks

The author says an international UMI data collection effort grew from five to 90 operators in eight weeks, built a dataset of over one million unique tasks and continues to capture over 45,000 new tasks daily. The author argues that robotic world models need much more task diversity than imitation learning. Pantheon, sharing the post, says its approach to robot foundation models has different requirements and calls its dataset one of the most diverse in robotics.