The speaker attributes some of the model's rapid task learning to pre-training on human data before the robot hardware was ready.

“the model has not seen this in pre-training in some other pre-trained model it's able to learn the task that you saw like very quickly and i credit a lot of that to like you know like even before the hardware is ready we're already committed to pre-training on humans because we knew at some point our robots will be like human you can get ahead on the training side while hardware is getting ready and i think now we're seeing that like the transfer is actually there and is actually very helpful”