The dataset carried 35 distinct task strings for 22 distinct tasks. The extra 13 were typos and phrasing variants that fragment language conditioning -- a VLA treats 'pick up the blue bird', 'pick up the blu bird' and 'ick up the blue bird' as three unrelated instructions. 'pcik up the batman toy and put it in the box' had split 4 episodes off the best-supported task.
Merged:
puth the blue bird on the table -> put the blue bird on the table
pick up the blu bird / ick up the blue bird -> pick up the blue bird
puth the batman toy in the box -> put the batman toy in the box
ick up the banana and put it in the box -> pick up the banana and put it in the box
pcik up the batman toy and put it in the box -> pick up the batman toy and put it in the box
put the bird in the box -> put the blue bird in the box
pick up the banana toy -> pick up the banana
put the banana on the box -> put…
