Google Wants to Solve Robotic Grasping by Letting Robots Learn for Themselves
spectrum.ieee.org
spectrum.ieee.org
BabyX First Words https://vimeo.com/103501130
Flexible Muscle-Based Locomotion for Bipedal Creatures https://vimeo.com/79098420
https://scholar.google.com/citations?view_op=view_citation&h...
A powerful and incensed AI could destroy the foundations of global society before we would even know what's going on.
A powerful and incensed President could order a nuclear strike that could end civilization as we know it. Why the concern over a theoretical intelligence that could theoretically have emotion and theoretically have so much power it could theoretically do the same?
It doesn't matter what we define as a soul, it is the relationship humanity will have with artificial intelligences that matters. If we don't ask these questions, our children will wish we had...if they're still around.
http://www.nybooks.com/articles/2015/10/08/amazing-inner-liv...
We can't know whether animals have a soul any more conclusively than we can know that about ourselves. But it's arguable that one is not less probable than the other.
In 'The Matrix' they had developed programs that could be uploaded to a physical brain to essentially pre-wire the brain with complex synaptic connections (or so I imagine was the effect). That would be a lot more efficient than waiting 3 years just to get to the point of trimming useless connections. We may have to 'grow' a virtual brain via training to develop the basic platform, but then pre-load it with operations to save time.
What does this even mean?
Add a clause for mostly being at least as capable of doing this as an average human, if you want to be sure. At the moment this is closer to a philosophical idea than an actual scientific possibility, but science usually makes its biggest advances at the edge of philosophy. This is also where we can get an idea of the high-level problems that need to be solved. Think the idea of manned space exploration before the advent of rocketry, for instance.
That is actually one of the more useful things a computer can do. It is also one of the hardest to program (see the halting problem). It is very useful for a computer for a computer to be able to do that when it has to pick out millions of likely candidates from quadrillions of possible answers.
>but they'll be just that—task solvers. Why make it more magical than that?
So you're in the 'people are magic' camp. You'll be very disappointed when you learn that people are just billions of years of bio-chemistry left to run rampant and that electro-mechanics and reproduce the same thing.
...how are you extrapolating that from anything I said? All I mean is that "AI" is inherently meaningless. When you ground it in reality there are more useful terms that are meaningful.
Also, I meant to point out that large parts of the brain's chemistry are not useful to emulate, leading me to question using humans as a base for "intelligence" comparison
AlphaGo, for example, can only play Go. It's never going to learn to play Poker, Scrabble, or Doom, without being completely reprogrammed.
It seems to come down to being able to interpret language in a way that is broadly within the range of a human's ability to agree with another human on the meaning of a phrase.
http://people.csail.mit.edu/brooks/papers/CMAA-group.pdf
If I may paraphase, his model is biologically inspired -- believing hierarchical layers of behaviours, lack of a central planning model (distributed processing), and physical and temporal placement in the world (rather than abstractions of the world, or observe/process/react loops) are essential to the formation of a truly intelligent machine.
The question of course, is, how could we deal with the implications of not just creating a smarter computer, or new life, but an entirely new and unprecedented class of life?
For one thing, if they're so great, why bother keeping us around?
If they somehow evolved past mechanical constructs towards more chemical/biological machines, or some sort of electro-mechanical construct that wasn't constrained by a physical/machine form, they would, as far as i'm aware, still require the biosphere to propagate or at the very least interact with us.
Even if they did want to fuel a spacefaring civilization (why???), why would they cannibalize the earth when they can already travel to planets, moons, asteroids, etc which have more of the components they actually need to function? Unlimited access to solar power, more heavy metals, and aren't restricted to the limitations and detriments of our biosphere (oxidation, gravity, ionosphere, electrically conductive material falling on you all the time, etc). On top of all that, there'd be these damn adaptable irrational homicidal monkeys that fear anything that could threaten them, and many things that don't. It would behoove them to get the hell away from Earth as fast as possible.
--
Anyway, all this assumes that somehow they could adapt faster than humans. I know of no plausible timeframe that anyone can yet estimate when this would happen. Kurzweil says 2045, but this makes no sense when you look at our actual progress over the past 50 years.
We still travel on Boeing 747s fueled by million-year-old decomposed plant juice. The richest nation in the world has one of the least advanced train systems in the world that only covers a fraction of the country. And one of the leading causes of death are the heavy speeding metal boxes we manually pilot through free space that are insanely less efficient than any other modern mode of transportation, purely because we like cars the most.
Predictions of technological advancement over 15 years in the future don't work because human progress is not based on capability. Much like your fear of a fate you can't be sure could even happen, our future is not based on what could happen, but what we make happen - our will.
Just as it is "possible" for us to birth a matricidal organism, it's just as possible for us to willingly never create such an organism. And since we imagine this supposed event would end in our own destruction, and we are [in general] on the side of self-preservation, it's not likely we would willingly pursue this event. It would be more likely if the idea never occurred to us.
--
Great civilizations have existed several times with advanced science, then fell, and took several millennia just for others to get back to where they had been. In terms of what is actually a threat to ours today, climate change and an unstable political/socioeconomic climate actually will kill us in a few decades if left unchecked. I suggest we focus on the immediate threats before we sweat the theoretical and philosophical ones.
Because we don't know the probability these things would be 'matricidal' does in no way justfity a 50-50 statistic. I'd say, we must be very, very careful not to. Its not at all clear what the long-term goals of a device designed to be our slave would be. Of course nobody would intentionally design it to be a threat to our civilization. The entire notion is, that our society is actually not considering long-term results of anything else; this would be just like all the rest of our inventions: agriculture, fracking, indiscriminate travel and population mixing (providing unlimited disease vectors), politics, money and on and on.
There are no trilobites around either.
The game engines don't do complex friction, as between gripper and target or foot and ground, very well.
The two main issues are that it is computationally expensive, but also that your mechanical modeling has to closely match that of the actual robot (especially the contact model), otherwise I suspect the training data will be useless in the end.
So if you can afford an actual robot, it makes sense to do the training using it.
On a typical desktop computer, that would probably range from real time for the fast/imprecise simulation, to maybe one day for a full-blown simulation.
But again, most roboticists will tell you there is a world between the simulation (even an accurate one) and the actual robot.
1. The best / most recent attempt at this was for the DARPA robotics challenge and the Gazebo simulator.
This was still very buggy and prone to hilarious / depressing physics.
2. Almost all game physics engines start from rigid body and slap on particles, deformables, etc.
An exciting counter example to this is nVidia Flex which starts with unified particle simulation (much closer to molecular dynamics simulation used for, you know, real work).
3. From the perspective of AI, accurate simulation might not be required.
Intelligence requires complexity and a certain degree of predictability. So as long as you can build a rich and consistent / learnable world then whatever simulation you have could be super useful.
From the perspective of transferring that knowledge into a robot though you need accurate physics.
4. Natural touch sensors are hard to do in rigid body simulators but are super important to naturalistic learning.
There's a ton of information that your sense of touch and body position provide about how the world works, and getting the tens of thousands of soft-contact touch points simulated you need for this kind of sensing is pretty challenging today.
Lots of physics engines do all sorts of things to minimize contact points, or ignore them if there's no motion. You have to work against optimization a lot if you want mechanoreceptors and proprioception.
I also saw a couple startups aimed at sub-$5k robots (like carbon.ai). Are they solving this problem in some novel way?
It's important for robots to not have backlash, because as movements are repeated, each bit of backlash adds up into a potentially big cumulative error. It could end up with the robot operating outside of the intended design envelope, which might be a safety problem.
I don't know what the startups are doing.
My guess is you can go pretty far with janky parts if you don't run for long periods of time and also measure where they are.
Their system evolves a virtual body which is evaluated by comparing its predicted behaviour (e.g. if motor A is rotated by X degrees, sensor B should get response Y) to real physical movements (moving motor A and reading sensor B). Once an accurate virtual body has been made, it's used to evaluate a bunch of (again, evolved) movement styles in simulation. Once an efficient style has been found, it's used to control the physical motors on the robot.
Also related, their lab has a "universal gripper" made out of a balloon filled with coffee granules: http://creativemachines.cornell.edu/positive_pressure_grippe...
In my case, I was working on biologically-inspired models for picking up distant objects. It's impractical to tune hyperparameters in hardware, so you need to be able to create a virtual version that gets you close enough. Once you can demonstrate success there, you then have to move to the physical robot, which introduces several additional challenges: 1) imperfections in your actual hardware behavior vs idealized simulated ones, 2) real-world sensor noise and constraints, 3) dealing with real-world timing and inputs instead of a clean, lock-step simulated environment, 4) having different API to poll sensors/actuate servos between virtual and hardware robots, and 5) ensuring that your trained model can be transferred effectively between your virtual and hardware robot control system.
I was able to solve these issues for my particular constrained research use case, and was pretty happy with the results. You can see a demo reel of the robot here: https://www.youtube.com/watch?v=EoIXFKVGaXw
Brett robot folding clothes. https://www.youtube.com/watch?v=Thpjk69h9P8
I can't recomend his book enough "Growing up with Lucy"