Status: No progress, just an idea. But man, I'd drop nearly anything if there was a shot to work on this.
SayCan really changed the way I looked at robotics — the robot didn't need to be trained on a whole, integrated process. Destructuring changes everything, and opens up the possibility of a robot with skills developed in a distributed fashion.
I propose a single hardware platform for robots, combined with a repository of atomic tasks, and a middle layer of a model that destructures every command into a sequence of atomic tasks.
There is no iPhone of robotics out yet — in fact, there doesn't even seem to be any sub $1000 robot arms designed for consumer use that aren't toys or "stem education tools".
A movable rig with a robot arm, a depth camera, a cheap tablet for display + user input, and a low powerded on board computer that can interface with a cloud instance. If we can build this at a $1000-2000 price point, consumer robotics starts getting a lot more interesting.
It feels like a huge number of tasks can be split up into further sub tasks, and eventually into atomic tasks. If thousands of hobbyists get the hardware platform we talk about, they can start developing these tasks and atoms.
These atoms and tasks will be interoperable: tasks can refer to and use other tasks as part of accomplishing their objective. All tasks are made/trained to work with the same hardware.
If we have a repository of all tasks available we can train/fine-tune a lagnuage model to split any instruction into these tasks, and make a task sequence. SayCan style :)
The start to this is by making a company for robotics hobbyists, and then there is a clear and obvious play for this company to be positioned incredibly to become the consumer robot co.