I guess they wanted to make robots less scary by rooting the story in the idea of babies. But it doesn't really achieve its purpose since there's still the obvious narrative of "oh, so what happends when the baby robots grows up into scary adult robots with all the greed and hate that that can be associated to". I think Googles Waymo took a better approach to framing the narrative PR around AI, namely making the robot feel like a toy or animation character instead.
Well, when the AIs are inevitably put in charge of a country's defense and another country is wiped out "by mistake" because of it, at least we can say the AI system that did it is a relatable cartoon villain.
Not it at all. I saw this paper presented at CoRL last month. It, and a lot of other papers on the topic of unsupervised learning in robotics, draw on ways humans develop cognition, specifically, during early childhood. In projecting the scene forward in the visual spatial domain, the robot plans it's actions and learns from the corresponding results in a way that parallels what a young child would do to learn about the world.
Eh, huge difference between "can't" and "won't". I feel confident in assuring you that they likely can imagine such futures, and likely do but with perhaps with extremely limited scope. The machinery is all there, and is probably working, even if they don't recognize it as such.
Also, like decision making in general, it's a skill that requires practice to improve.
This is cool research, but it should be noted that it's a pretty basic extension of prior work on video prediction:
"the robots can predict what their cameras will see if they perform a particular sequence of movements. These robotic imaginations are still relatively simple for now – predictions made only several seconds into the future – but they are enough for the robot to figure out how to move objects around on a table without disturbing obstacles"
There is no effort at all at modeling the world at a more abstract conceptual level, so pretty fundamental ideas/changes will be needed in order to do 'long-horizon' planning/imagination.
Finally! Common sense is being able to predict the consequences of your actions. I've wanted to do that for years, but I was thinking in terms of planning and trial and error against low-resolution physics models of the world. Sort of like A* on steroids. This has limited applicability, although some video games try to do something like that. The new approach here is more promising, because it needs less of a world model.
Getting through the next few seconds without a major screwup is the big problem in unstructured robotics. This approach may not be the right answer, but it's addressing the right question.
11 comments
[ 4.6 ms ] story [ 34.0 ms ] threadAlso it looks like the relevant paper is here[3] and a video example is here[4].
[0]: http://slatestarcodex.com/2017/09/05/book-review-surfing-unc...
[1]: https://www.amazon.com/gp/product/0190217014/
[3]: https://arxiv.org/abs/1610.00696
[4]: https://sites.google.com/site/robotforesight/
Some AI literature is of the belief that consciousness emerges from imagining hypotheticals. You could probably describe a few key milestones.
1 - Imagine your hypothetical actions, interacting with objects.
..
.. - Imagine the actions of other 'entities' capable of imagining actions, who are also interacting with objects.
..
.. - Imagine your 'self' in relation to these other entities, all capable of imagining each other's actions.
Also, like decision making in general, it's a skill that requires practice to improve.
"the robots can predict what their cameras will see if they perform a particular sequence of movements. These robotic imaginations are still relatively simple for now – predictions made only several seconds into the future – but they are enough for the robot to figure out how to move objects around on a table without disturbing obstacles"
There is no effort at all at modeling the world at a more abstract conceptual level, so pretty fundamental ideas/changes will be needed in order to do 'long-horizon' planning/imagination.
Getting through the next few seconds without a major screwup is the big problem in unstructured robotics. This approach may not be the right answer, but it's addressing the right question.