You'll see many amazing robots doing amazing demos
but there's no robot in the world that can come to your messy apartment clean your kitchen and wash
all your dirty clothes and clean it all up. There's just no robot that can handle
all those different types of things I'm Pete Florence I'm Lucas Manually
We're here in the robot locomotion group at MIT.
Tthis latest work is self supervised correspondence in motor policy learning.
I think what's particularly exciting about it is that we can enable
robots to do a wide variety of tasks with a very small amount of human effort
for example for some of these tasks we used as few as just six minutes of
training data for the robot to then be able to do the task reliably.
So we want to put in a very small amount of human effort and have the robot learn a new
task such that we can put in a small amount of human effort for one task and
another task and eventually be able to scale up such that robots can do a whole
huge variety of helpful tasks. This is a quite different paradigm to be in
In a lot of traditional robotics where the robot basically looks once at
the environment and then decides what to do and closes its eyes that's much
different than constantly looking at what's happening in the world and
constantly deciding new things in the second paradigm where we're doing this
real-time feedback it certainly helps them handle the wide diversity of
objects and physics that robots will encounter in the world