News for hackers. Stuff that matters.
Re:Building a deeper understanding of images by mrfusion on Friday September 05, 2014 @12:35PM (#8274805)
That's pretty amazing. It seems like we're at a point where we could build really practical robots with this?

Robots to do dishes, weed crops, pick fruit? Why isn't this being applied to more tasks?

Re:Building a deeper understanding of images by karpathy on Friday September 05, 2014 @02:11PM (#8275350)
I think one of the main reasons is that the improvements have been drastically significant and very recent; There hasn't been enough time to convert the research code into open source code and libraries, but you can confidently expect these models to become pervasive over the next few years not just in robotics, but in all perception systems.

A few good open source libraries out there that have the building blocks for putting together similar models: - Caffe (C++) http://caffe.berkeleyvision.org/ - cudaconvnet2 (C++) https://code.google.com/p/cuda-convnet2/ - DeepBeliefSDK (iOS,Android,OS X, Raspberry Pi, JS) https://github.com/jetpacapp/DeepBeliefSDK

Re:Building a deeper understanding of images by mdda on Friday September 05, 2014 @03:54PM (#8275948)
karpathy is being too modest here. He's also created his own convnet.js [1], which you can play with online [2] (complete with relevant, working demos), etc.

[1] https://github.com/karpathy/convnetjs [2] http://cs.stanford.edu/people/karpathy/convnetjs/

Re:Building a deeper understanding of images by drcode on Friday September 05, 2014 @12:46PM (#8274865)
I would eat a "hat with a wide brim" if Google isn't going to release a robot that can do basic household chores (laundry, dishes, dusting) within the next 3 years.

Google has been gobbling up robotics startups, and given how Google also loves gobbling up personal data, having robot "boots on the ground" in every home must be extremely appealing to them.

Re:Building a deeper understanding of images by jrussino on Friday September 05, 2014 @04:25PM (#8276088)
The idea is right but if by "release" you mean "sell to consumers" I think you're a little over-optimistic about the timeline.

Google's driverless car project is a direct descendant of the tech developed for the DARPA Grand/Urban Challenges. Those took place in 2005-2007 [1]. It's been ~9 years since the first Grand Challenge, and while there has been great progress in this area no one will sell me an autonomous car just yet.

The first DARPA robotics challenge was held in December of 2013 [2]. If you check out some of the videos from that competition [3] [4] (or some of the PR2 videos another user posted), you'll see that we can make general-purpose humanoids that can do some pretty neat stuff, but there's still a lot of work to be done before I can buy one that will reliably do many different household tasks.

The Atlas platform that many of the competitors used in the DRC was developed by Boston Dynamics, who were subsequently bought by Google. So yes, they are definitely in this space. Maybe progress in humanoids will go faster just by virtue of the fact that we've got an extra decade of research to build on, but the journey from "works in a lab/demo, barely/slowly" to a viable commercial product is a long one.

[1] http://en.wikipedia.org/wiki/DARPA_Grand_Challenge [2] http://en.wikipedia.org/wiki/DARPA_Robotics_Challenge [3] http://www.youtube.com/watch?v=hzmMVHGNXvI (sped-up highlights) [4] http://www.youtube.com/watch?v=mwWm3HaDbnQ (full 10-hour recording of competition live-stream)

Re:Building a deeper understanding of images by VikingCoder on Friday September 05, 2014 @01:40PM (#8275152)
I'll congratulate you on not specifying a time duration for your meal. Indigestible material can cause a blockage in your colon, which can cause severe pain, damage to your colon, and even death. I suggest chopping the hat up into very, very small pieces, and eating them over a very long period, such as a month or longer.
Re:Building a deeper understanding of images by jschwartzi on Saturday September 06, 2014 @02:13PM (#8279010)
If you'll excuse me, I have to write some letters to the legal department at the hat factory where I work.
Re:Building a deeper understanding of images by robotresearcher on Friday September 05, 2014 @02:17PM (#8275389)
I doubt it. The easy parts of laundry and dishes are already done by simple robots sold in every white goods department. The remaining parts are very demanding indeed.

Research labs are not robustly demonstrating these capabilities yet, even with very expensive robots.

Re:Building a deeper understanding of images by botman on Friday September 05, 2014 @05:17PM (#8276353)
Another robot researcher agrees. Classifying images is totally different from enabling a robot to perform a difficult task. Object recognition is a supervised learning problem, mapping image -> vector of probabilities. Robotic manipulation is a control problem, mapping a long sequence of images -> a long sequence of actions.
Re:Building a deeper understanding of images by bfung on Friday September 05, 2014 @03:31PM (#8275838)
I had this thought the other day - not robots washing the laundry, but folding it. Folding is actually not that simple after thinking about it.
Re:Building a deeper understanding of images by robotresearcher on Friday September 05, 2014 @04:05PM (#8275992)
https://www.youtube.com/watch?v=gy5g33S0Gzo

Work is progressing. That's a $300K robot.

(edit: no affiliation. Video shows PR2 robot at Berkeley folding towels competently but very slowly in 2010)

Re:Building a deeper understanding of images by GhotiFish on Friday September 05, 2014 @01:02PM (#8274959)
... Do you want me to note that in my calender?
Re:Building a deeper understanding of images by HSO on Friday September 05, 2014 @01:05PM (#8274977)
Re:Building a deeper understanding of images by robotresearcher on Friday September 05, 2014 @02:07PM (#8275323)
Robots are being applied as fast as (i) they really work, and (ii) they make economic sense.

Precise, powerful mechanical actuation is really expensive. High quality motors use expensive rare-earth magnets and require very accurate manufacturing.

Right now, only very valuable manipulation tasks can justify using a robot. Hence, robots built most of my car, but I take out my own trash.

Non-physical AI stuff will catch on much faster than robots.

edit: The exception is clever things like Roomba, where the mechanical parts are dirt cheap, and the robot's behaviour compensates for its lack of precision. Neat.

Re:Building a deeper understanding of images by krasin on Friday September 05, 2014 @04:16PM (#8276052)
I only partially agree. Just recently, pretty decent stepper motors have become very cheap (~$6 for NEMA-17), thx to all this 3d printer boom, and lots of things are now finally possible to implement on a budget.

But I agree that good servos with high torque are still super-expensive.

Re:Building a deeper understanding of images by robotresearcher on Friday September 05, 2014 @04:57PM (#8276245)
Some manipulation tasks require force control, too.
Re:Building a deeper understanding of images by krasin on Friday September 05, 2014 @06:42PM (#8276695)
Surprisingly, this is also getting solved. For example, there're demos of self-calibrating cheap Delta robots which use force sensors on the plate: https://groups.google.com/forum/#!topic/deltabot/6fxnM20nYKc

Sure, it's not yet there, but with the space of feasible is growing fast.

Re:Building a deeper understanding of images by robotresearcher on Sunday September 07, 2014 @02:07PM (#8281885)
I'm not an expert on 3D printing, but it looks like this uses force sensing to make sure the effector velocity vector is nicely parallel to the position of the base plate. The effector is not exerting any force on an object. So it's velocity control not (the more demanding) force control.

Let me know if I have misunderstood what's happening here.

Re:Building a deeper understanding of images by krasin on Monday September 08, 2014 @11:49AM (#8285418)
You're right, I was slightly off in my example.
Re:Building a deeper understanding of images by Houshalter on Friday September 05, 2014 @03:50PM (#8275935)
I wonder if using imprecise motors would be acceptable with sufficient AI. So it could account for errors and imprecision in it's movement.
Re:Building a deeper understanding of images by robotresearcher on Friday September 05, 2014 @04:08PM (#8276008)
Example project partly motivated by that idea:

http://cswww.essex.ac.uk/staff/owen/machine/cronos.html

(no affiliation)

Re:Building a deeper understanding of images by michaelbuckbee on Friday September 05, 2014 @01:47PM (#8275180)
We don't really classify it this way, but Google's self driving car is the practical robotic implementation of this. There has been a bunch of press around how the Google self-driving car has to have streets mapped out for it, etc. This research will go directly to issues like: "Is that a 'domestic cat' or 'paper bag' in the street up ahead?"
Re:Building a deeper understanding of images by jschwartzi on Saturday September 06, 2014 @05:19PM (#8279522)
What if it's a domestic cat in a paper bag?
Re:Building a deeper understanding of images by sp332 on Friday September 05, 2014 @12:42PM (#8274840)
Because it was just invented over the past 2 years? And it is being used, but mostly by Google and Facebook.
Re:Building a deeper understanding of images by imaginenore on Friday September 05, 2014 @02:12PM (#8275359)
Re:Building a deeper understanding of images by cscurmudgeon on Friday September 05, 2014 @12:49PM (#8274878)
Just a small list of other things you need:

1. Natural language understanding

2. 3D object understanding

3. Planning

4. Object manipulation

5. Navigation

6. Speech recognition

Re:Building a deeper understanding of images by tachyonbeam on Friday September 05, 2014 @12:59PM (#8274941)
You can realistically skip 1 and 6, control it with your smartphone instead. Or, you can simplify the problem by using more constrained verbal commands. I was told that speech recognition works very well when you have a simple grammar instead of an open-ended language.
Re:Building a deeper understanding of images by mrfusion on Friday September 05, 2014 @12:56PM (#8274924)
Really, just for pulling weeds? Maybe you thought I meant a robot that can do arbitrary household chores. I actually meant specialized robots for specific tasks.
Re:Building a deeper understanding of images by icegreentea on Friday September 05, 2014 @01:52PM (#8275224)
You skip 1 and 6 then. 3 and 5 isn't bad right now. 4 might be tricky (dependent on task at hand - weed pulling can be surprisingly difficult depending on context). 2 would likely be difficult.

The results for the actual competition are here. You can take a quick skip at the error rates of the first place approaches. For some tasks they're down to 5-10% which might be acceptable for some tasks, but could also be completely unacceptable for others. shrug

The day is probably coming, but it probably won't be soon. Not the least because robots are actually ridiculous to work with / develop.

http://www.image-net.org/challenges/LSVRC/2014/results