I propose a grand challenge for dextrous robotic manipulation: cook eggs in all the ways shown in this videohttps://www.youtube.com/watch?v=BW1x6F63F80 …
Prikaži ovu nit
-
3/ Cooking eggs is very timing-sensitive - forces robots to be quick. Slow robots (academic manipulation research does not optimize for this) would fail instantly.
Prikaži ovu nit
4/ current computer vision algorithms have a very pose-centric, object way of understanding scenes. Having to manipulate an amorphous scramble in a pan or watch a poached egg disappear under a swirling mass of egg white would likely break those abstractions.
-
-
5/ it's still hard to imagine general purpose robots with our current software and hardware, i.e. defining useful tasks for robots without solving AGI. I think an "eggs any style" robot would be a concretely useful, well-defined problem scope with real utility to real people.
Prikaži ovu nit -
6/ Many objects in human spaces have functional properties (doors open, hand tools convert energy, buttons are pressed). An egg is not typically thought of as a tool, but it has as many functional forms as there are ways to cook it. Great object to master manipulation skill over.
Prikaži ovu nit -
7/ For those interested in RL research, maybe instead of trying to estimate empowerment of a policy in an entire world, you can more easily estimate the empowerment w.r.t. a single versatile object (the egg).
Prikaži ovu nit -
8/ Simulations of reality often lack the functional properties of modeled objects (never mind chemical properties). Cooking eggs would be a good forcing function to encourage us to not overfit to or rely to heavily on simulation.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.