How transferable do you think the model trained for one robot will be to others? At the high-level of task planning this is obviously the case, but that piece is in some ways also the easiest to already accomplish using model-based planning. So I'm more interested in the "difficult" part of manipulation like grasping. Unlike words, where the same word is the same word, picking up the same object with a different robot is not the same. Could that jeopardize the idea that 2T tokens is enough and/or the idea that foundational robot models are even transferable, meaning that in fact each robot (type) would need to collect its own 2T tokens?
I think transfer across robots will follow the same "scaling laws" as everything else, which is to say that with sufficient diversity of embodiment, sampled along a good distribution, it will be fine. So, with 100 robots evenly sampled in robot space, generalizing to robot 101 will be easy. The problem is that this just isn't how robot data works right now.
How transferable do you think the model trained for one robot will be to others? At the high-level of task planning this is obviously the case, but that piece is in some ways also the easiest to already accomplish using model-based planning. So I'm more interested in the "difficult" part of manipulation like grasping. Unlike words, where the same word is the same word, picking up the same object with a different robot is not the same. Could that jeopardize the idea that 2T tokens is enough and/or the idea that foundational robot models are even transferable, meaning that in fact each robot (type) would need to collect its own 2T tokens?
I think transfer across robots will follow the same "scaling laws" as everything else, which is to say that with sufficient diversity of embodiment, sampled along a good distribution, it will be fine. So, with 100 robots evenly sampled in robot space, generalizing to robot 101 will be easy. The problem is that this just isn't how robot data works right now.