Cognitive Robotics [Discussion]

Perhaps we are at the age of the turning point in artificial intelligence. Perhaps tomorrow one of us on the planet will prove that machine ALREADY has consciousness and all human will recognize them as LIVING BEINGS! ohohohoho

What do you think about cognitive robots? Are they close to a life form or not?

No, they can't even make me a coffee. How can they even be recognized as living beings?

Actually one robot can: the Robobarista! Its goal is to tackle manipulation of novel objects by directly transferring a manipulation trajectory from a previously manipulated object by mapping visual data, human language instructions and trajectories onto a lower space. They called their approach deep multimodal embedding, and used crowdsourcing to obtain training data for their neuro net.

LOL it taught me that "the motion required to operate the handle of the espresso machine is almost identical to the motion required to flush a urinal with a handle".

They are using part-based transfer, so of course the handle part of the espresso machine look a lot like the handle of a urinal flusher. Now everyone can think of this next time they make an espresso, cappuccino or latte.

Bu the accuracy of this method is quite low because it's a direct transfer and it's based on local parts. Maybe extending it to adapt selected trajectory by recognizing feedbacks and also have a more global understanding of the environment can be useful.

Yeah, the jellyfish is right about accuracy. See robot can't become as smart as a barista because they suck at the global cases. We can only make them do things in a very constrained world.

Nonetheless, robots have made a large improvement, especially with neuro nets. We can see that pattern recognition has made a huge quantum leap. AlphaGo has defeated the best world Go player just now. I think soon, we will be able to deal with more and more general cases. It won't take long for neuroscience to understand the brain enough for roboticists to replicate it.

Building a robot, the ideal humanoid that people are imagining, is quite a challenge on many levels. It does not merely involve high-tech and engineering skills. It poses questions related to biology, neurology and even ethic and philosophy. Because we are basically building another moving entity with the ability to interact with ourselves and this world. It's no longer an isolated system in the laboratory or on the Web. It's really another being, with possibly infinite potential.

So that's why it's hard to build such robot, and up to now, the robot evolution is at a premature stage. So of course, when you look at something like Robobarista, it looks stupid on many level, and its video demonstration seems to have cut off many (probably embarrassing) things.

Spirit wrote:Nonetheless, robots have made a large improvement, especially with neuro nets. We can see that pattern recognition has made a huge quantum leap. AlphaGo has defeated the best world Go player just now. I think soon, we will be able to deal with more and more general cases. It won't take long for neuroscience to understand the brain enough for roboticists to replicate it.

I think what we lack right now is the understanding of how neuro network get good results for pattern recognition and how those learned features in the layers corresponds to who human perceive the world.

shadow wrote:Building a robot, the ideal humanoid that people are imagining, is quite a challenge on many levels. It does not merely involve high-tech and engineering skills. It poses questions related to biology, neurology and even ethic and philosophy. Because we are basically building another moving entity with the ability to interact with ourselves and this world. It's no longer an isolated system in the laboratory or on the Web. It's really another being, with possibly infinite potential.

I agree that robotic includes many other fields and I think it's a shame that people in our lab are usually just engineers, computer scientist, mathematician, etc. Basically they are all in the science or engineering fields. There's no people in arts, history, humanity, philosophy, finance, etc.

I just read the paper "Language within our Grasp" (link) about neurons that recognizes actions, and I think the approach in Robobarista can be an analogy to this neurological phenomenon. The paper talks about an area in the brain that is triggered when observing a certain action and when doing that same action. At some level, this sounds similar to a neuro network that processes action trajectories. Another thing that I find the most interesting is that the paper talks about how the brain area responsible for mimicking actions might be the ancestor of the brain area responsible for speech. This raise the hypothesis that language (or simply labels) originated from the need of recognizing and classifying actions for the purpose of mimicry, which represents the method for sharing information and knowledge.