Robot With Broken Leg Learns To Walk Again In 2 Minutes | Hacker News Reader

Robot With Broken Leg Learns To Walk Again In 2 Minutes

medium.com

75 pointsspectruman30 comments

The "ground contact" parameter while walking sounds functionally very close to the "proprioception + pain" parameters for animals in a hobbling situation. An animal (me, specifically) quickly discovers a hobbling strategy by experimentation and a continuous sense of pain levels.

I think that pain will become a major part of the design in learning robots of the future.

I think that pain will become a major part of the design in learning robots of the future.

This is something that Hollywood has been warning us about. We do not want to program robots to "feel" pain. We should build in avoidance algorithms, but not pain.

The problem is that anything that is "aversive" in the sense of avoidance may ultimately be perceived as pain. Whether we need to associate a emotional valence with pain in the way evolution has (because evolution coopted existing systems) is an open question. We may not need to, but any function that takes multiple pain like inputs and associates with with a conserved set of behavioral responses is likely to result in something similar to our perceived experience of pain.

Hollywood? :-(

I believe fiction plays an important role in shaping the philosophy of science. Forward-looking science fiction, especially, considers issues that have yet to be encountered by society, providing the opportunity to decide how to handle or avoid those issues in advance.

In the case of robots, I was imagining a line from the less-than-serious The Simpsons, when a robot flees a flaming building, asking, "Why was I programmed to feel pain?" ( http://m.youtube.com/watch?v=nQ-ggzfdsMs ). Indeed, why would a robot be designed to suffer when enough suffering exists already? We should design our non-sentient robots now with that thought in mind, in anticipation of the possibility of creating self-aware AI.

Feeling "pain" is not the problem. The displeasure experienced due to the pain is.

Part of me loves this and the rest of me worries about how little chance I'll stand surviving the robot apocalypse. The adapted behaviour looks like a wounded, yet unstoppable, machine hell-bent on carrying out its mission.

In a different life I'd definitely have worked in robotics.

Well, I think we'll merge humans and robots before having advanced AI robots.

Basically cyborgs will be the ultimate creation, and since we'll have a human brain with empathy (hopefully) then we will be able to prevent any uprising before it happens.

Just look at all the advanced stuff that DARPA is funding and succeeding in creating. In the next 20 years there is going to be some insanely awesome technology coming out.

I don't like the idea of creating 2 classes of humans though, (those with BCIs/other advanced technology embedded in them and those without) that just seems like it's going to cause a problem.

It would be really nice to have an alien civilization come down and tell us the potential mistakes to try to avoid when maturing. Technology is so close to increasing at an exponential rate that something is almost certainly going to wipe us all off the planet. We'll be very lucky if we don't end up killing ourselves in the next 50 - 100 years, in my opinion of course.

A lot of good human empathy has done the human race... I think i'd rather trust robot empathy.

"Well, I think we'll merge humans and robots before having advanced AI robots."

That seems like extremely wishful thinking to me.

Anecdotally, it looks like progress in making machines do human like things is progressing much more quickly than reverse engineering the human brain.

He doesn't mean human-like robots (with a human-like brain). He means a human brain with partly machine body.

Even that's real damn hard, though, and I'm not optimistic.

I thought good progress was already being made with this, with devices like cochlear implants, retinal implants, and cyborg monkey arms: https://www.youtube.com/watch?v=wxIgdOlT2cY

Yep, lots of teams working on BCIs.

The rapid learning by introducing an electric current at a specific certain parts of the brain: http://en.wikipedia.org/wiki/Transcranial_direct-current_sti...

http://www.technologyreview.com/view/427177/diy-kit-overcloc...

BCIs are getting more and more attention. Imagine being able to hook up a processor, external storage, or even directly connecting to the internet using a BCI. Even connecting two brain is possible (and has been done). Some people say that one day we'll all be one connected organism, when we die just throw the brains in a big tank and let them live forever, like Futurama.

http://www.technologyreview.com/news/527561/military-funds-b...

http://www.army-technology.com/features/featurebrain-compute...

It seems very accurate to me, when we consider how many non-permanently attached machines we already use. A smart phone, the Internet, search engines, prosthetic limbs, and cars would certainly allow for a powerful cyborg. Is the only major distinction that these are not permanently attached? Or does a cyborg also require neural interfaces?

Don't worry, the legs just weren't broken enough.

My 8th grade science fair project studied a very simple version of this problem! I used genetic algorithms to have a 6-legged robot relearn to walk after 1 leg was injured. It won 1st grand prize in Texas in 2004 :)

How is that a very simple version of the problem :)

Its more of a 'brute force' method ... genetic algorithms take one or more variables and randomly change them. Then keep iterating on a solution that works - you don't need to know why something works only that it works better than the previous solution.

While genetic algorithms are highly iterative and often computationally expensive approaches to randomized search, they are vastly more efficient actual brute force methods. It would be better to say "it is more of a 'randomized search' method", especially with the connotations associated to actual brute force techniques.

https://en.wikipedia.org/wiki/Random_search

The robot hardware was the very simple part. Each leg only had 1 degree of freedom and was binary, just contracted or extended with no in-between. It was the Stiquito robot http://en.wikipedia.org/wiki/Stiquito , which is dirt cheap and uses muscle wire (nitinol) to move the legs.

Reminds of these cool evolutionary algorithm simulations:

http://vimeo.com/79098420

Is it just me or does it look a lot more animal-like when it's limping?

Yeah, it's "pre-damage" gait looks really stiff. As if it's been pre-programmed, but not been subject to whatever learning algorthims the researchers have given it to "recover" with.

I think I'd like to see the robot move, after those algorithms have had a chance to improve it's initial non-damaged gait.

It looks like they did try that, and it moved faster:

> “When the robot is undamaged, our approach yields dynamic gaits that are 30% faster than the classic reference gait,” say Cully and co.

However, there don't appear to be videos of that.

The second video gave me an idea: it would be really great if the thing was damaged and could create a model of its damaged self in a simulation... Then it could play out different gaits (presumably at faster-than-reality speeds) to determine an optimal new gait. It doesn't seem like this team is too far away from that.

It would take some exercise of movement to quantify the damage and the resultant qualities.

Actually I believe that's exactly what the system does.

Okay... but actual animals do way better:

http://www.ted.com/talks/robert_full_the_secrets_of_nature_s...

As a rule of thumb, any walking robot that uses servos can't approach the performance of its animal colleagues. The leg impedance is way too high.

I love the part where the sim flips on its back and starts flailing. Thinking outside the box!

Url changed from http://spectrum.ieee.org/automaton/robotics/robotics-hardwar..., which points to this.

Reply on news.ycombinator.com