Figure just highlighted fingertip design as one of their major improvements, and it’s a massively underrated & overlooked part for robotic hands. Im happy they (and others) are now calling it out. Just because something “looks like” a hand/finger/finger tip doesnt mean it will perform like one You’ll see demos of hands picking up larger objects or being directly handed tools to manipulate in-hand, but rarely see end-to-end picking something tiny/thin off of a table and using it That capability comes from the fingertip, and fingertip design isnt easy. The human finger is so god damn impressive, and it’s impossible to replicate 100% (nonlinear deformities, durometer differences, skin texture/frictional changes, touch sensor capabilities/restrictions) so you end up needing to adjust the design to match capabilities. If you put a glove on yourself (literally any glove) and try to do tasks with smaller/thinner objects, youll notice a significant performance change compared to your own hand and you need to adjust your approach accordingly. Maybe thats why the figure demo had it picking up stacked plates a little differently than how we would do it as humans (or it could have just been an easier approach for a teleop user/sim to learn) You have 3 levers to pull for finger tips: shape, compliance, and surface finish And the touch sensors you choose automatically impose design constraints on some of these levers. Depending on your benchmarks, youll start to shift those levers - which can start to lead you away from exact human fingertip shape It’s interesting looking at the fingertips of hands that come out to see what changes they made to improve capabilities or if their only benchmark was “make look human”. You can really tell just from the finger tips what the robot is going to be useful for.

You can tell how useful a robot will be by the hands in which it yields