Waldoing is also a great way to raise funding. Robots are dumb and break during demos, so it's better to have a human in the loop. You demonstrate the *potential* of the hardware by offloading the intelligence to your local intern. This speaks to a deeper truth ... 2/n
-
-
Prikaži ovu nit
-
In many cases our hardware has been good enough for 30+ years. Specifically it has been good enough to allow a human to operate it and impress other humans. Humans are ridiculously good at manipulation, even when we're using bits of underpowered metal with substantial lag. 3/n
Prikaži ovu nit -
Consider the art of animatronic puppetry. Humans remap their hand movements to control the *faces* of creatures that have never existed. (See Jim Henson's Creature Shop which is truly a wondrous place https://youtu.be/WQyWTivri1g?t=461 …) 4/n
Prikaži ovu nit -
If a human is asked to control something that is vaguely shaped like a hand and can grasp things, it's child's play. Recall that we're blobs of jello-like goo in a bone box that have to learn to control two 27 degree of freedom manipulators to do things like "pick up food" 5/n
Prikaži ovu nit -
How does this apply to the subject of the paper? First, the expectations of other humans when it comes to manipulation are *insane*. If you show them an arm-like thing attempting an unconstrained task, the bar for "impressive" is very very high 6/n
Prikaži ovu nit -
Second, we know that economic value is largely about speeding processes by removing humans from the loop. Computers do math really fast, draw pictures really fast, and find information really fast, so we like them. So how do we apply this? 7/n
Prikaži ovu nit -
By accepting that human level performance is the *starting point* of expectations we must find applications where we can sell robots controlled by humans today in the expectation that we will be able to rapidly deploy advances in technology tomorrow. 8/n
Prikaži ovu nit -
The ideal place for this is not in homes or research labs, but in assembly line work cells. If your hardware + human can do the job slowly today (it will be slower) then tomorrow it can do the job slowly at 1/100th the hourly cost. 9/n
Prikaži ovu nit -
@kindredai gets this. The "work cell" in which they deploy is overly constrained unfortunately because they are trying to sell automation with today's technology. 10/nPrikaži ovu nit -
The other thing we need to recognize though is that optimizing for a single task is a *dead end*. Remember, my hands can juggle, cook, control the face of a dinosaur and type. We should recognize this as fine-tuning a general system and not learning a task from scratch. 11/n
Prikaži ovu nit -
As also pointed out in this issue of ImportAI, simulation is key to tractable data collection for real world tasks. We need to train flexible, very high DOF agents in a complex simulated world using a general objective like survival to produce the image-net moment we seek. 12/n
Prikaži ovu nit -
While it may not be strictly necessary I also believe the rapid *change* in our morphology as we grow is a regularizing force that will help our virtual agents generalize. 13/n
Prikaži ovu nit -
By building out a network of hardware that's doing real work today, even if it's just a proxy for humans, and training virtual agents with the broadest possible flexibility in minds, we can create the conditions for rapid progress, value creation, and deployment. Fin.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
Prikaži još odgovora, uključujući one koji mogu sadržavati uvredljiv sadržaj
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.