Legged Robots - Computerphile

How do you make a legged robot robust when the environment is against it? Ioannis Havoutis of Oxford Robotics Institute walks us through it.
/ computerphile
/ computer_phile
This video was filmed and edited by Sean Riley.
Computer Science at the University of Nottingham: bit.ly/nottscomputer
Computerphile is a sister project to Brady Haran's Numberphile. More at www.bradyharan.com

[@hpekristiansen]

9:11 Did you see that thing!? -so human like

coyotes are well known to communicate mainly with written signs on a pole.

[@Computerphile]

I wish I'd thought to include that in the animation, or at least a painting on a tunnel :) -Sean

[@Att4ni]

I just wanna appreciate how much effort goes into making and editing these videos and pushing out regular uploads. This was apparently shot on July 13 (8:50). Really appreciate the content and really great video as always!

[@octopuscabbage8091]

the funny thing about reinforcement learning people is they say the advantage of rl over mpc is "you don't need a model of your dynamics" but then immediately say "we train it in a simulator [with a model of the dynamics]"

[@fletcherreder6091]

Exactly! The dynamics are for the simulation people to worry about. Assuming you didn't write your own simulation, and if you have money for real hardware then you probably didn't.

[@oldcowbb]

whats your point, you train it with simulator for convenience, you still have to train it in real life afterward, point being the algorithm it self does not require model

[@woulfhound]

@@oldcowbb The point is that those people are hypocrites for saying they don't need a model of dynamics.

[@thekaxmax]

@@woulfhound 'doesn't need' is not the same as 'doesn't need but uses one anyway for convenience and speed'. Not hypocrites.

[@prakharmishra6331]

Loving computerphile's foray into robotics!

[@Pinefenario]

Nice video! I think with robotics we are at same point as with computers back in the eighthees.

[@fuzzynine]

Beautiful.

[@Olson2BW]

One thing I've always noticed is that the robots always seem to have the same power to the actuators all the way through the motion. You can take your arms and swing the lower part around without using any muscles in your forearm by hanging it down and using your upper arm to swing it back and forth. I have never seen a robot where you can do that. Why is this important? It is important because it wastes a lot of effort and energy. It also causes robots to be stiff and jerky compared to humans or animals where you can describe a lot of people and animals and graceful. When is the last time (the answer is never for me) where you saw a robot be graceful? I think that is one of the biggest reasons why robots will not "seem" like humans anytime soon unless they stop powering every second of an action. You do not strain your muscles unless you are lifting or moving a heavy load. Robots always seem like like they are using maximum effort for everything they do. If we did that with our bodies we would be exhausted and need to lay down most of the day or we would develop much bigger or stronger muscles with more endurance. Instead our brains have learned that you don't need maximum effort all of the time. In fact, --most-- of the time you do NOT need maximum effort. Instead you most often need less than half and in many cases less than 25%. I'd love to hear thoughts on this about potential smoothness of robots so that they would be a LOT more graceful AND for energy savings that could be had in this area.

[@humbertobrito8478]

I agree, very interesting point! I can imagine engineers will find a way to improve robots graceful movement. Although I feel uncanny about it, would be quite a spectacle. Imagine what this could do for the entertainment industry and practical effects. A giant robot spider or T-rex!

[@johnnybravo964]

Yes, I think if they train the robot to limit energy usage and also train to limit the force on the actuators then a robot will become more graceful and walk more naturaly

[@damianwiecaw492]

Not true, motors used in robots can generate torque proportional to current which can be any current value. The problem lays into control alghoritms because its very hard to implement

[@rvignesh2809]

Check out experiments by Prof Deepak Pathak at cmu

[@thekaxmax]

That's because the actuators are not muscles and give total control at every point. If we had different actuators than rotating geared motors you could do it differently. But there isn't anything else that's a useful size and power.

[@JuanPabloCarbajal]

Nice to see Ubuntu desktop in that screen cast

[@alexgeo1159]

Nice exhibition, συγχαρητήρια :)

[@JakeAikens]

Remember this video when one of these is chasing your family down the street.

[@thekaxmax]

although if that doesn't happen blame a politician--who will have made the decision to make them aggressive--rather than the roboticists.

[@accountname1047]

This guy is working on legged robots and zuckerberg can't even get legged pixels to work on a computer

[@aravindpallippara1577]

Cgi is often harder than real life lol

[@daveduncan8090]

Has anyone tried to gather all the information from CGI data mapping human motion to use for creating algorithms for models of robotic motion?

[@octopuscabbage8091]

Yeah that's called imitation learning. It's not quite that but it's very similar

[@thekaxmax]

yes, a lot, to determine how the skeleton does its thing. Mapping from one to the other is hard because the actuators are so different.

[@user-py7qh8bn8s]

leg moment

[@speedfastman]

We're getting closer and closer to robot humanoid cat girls. Hang on tight boys it's just around the corner.

[@MrIneedalifenow]

I love you

[@SAHarry]

Ew🤢

[@Reach41]

We’ll just wind up with another source of complaints.

[@sevdev9844]

They're already on the market or you need to wait a while for walking ones, dependent on the details of what you want. Walking and having constraints like that is tough.

[@SteveGouldinSpain]

Not FSD yet then.

[@tramsgar]

Surprise NDA ending?

[@george27121984]

Έλληνες παντού!!!

[@bitcrawl]

🗿

[@AcornElectron]

Neural networks running potentially self replicating robots… what could possibly go wrong?

[@audreyjuarez9450]

pr໐๓໐Ş๓

[@BD90..]

Looks like you use Roblox to make models lol

[@scowell]

For ongoing learning you'll have to have it 'bruise' its brain! Every time it falls some random number gets installed... that's how neural networks learn... there has to be pain (or death!).

[@oldcowbb]

the "pain" is encoded as negative reward

[@SharhbiniRauf]

You must put parendev motor + alternator, so your robot alive 50 years non stop.

[@AleksyGrabovski]

It's already 2022 and robots are still so lame...

[@gasdive]

4 leg robots seems to be a mistake. 4 legs are an accident of history, and never came from walking anyway. It seems that if you've got a many legs as you want, 8 is near optimum. Everything evolves into a crab. Watch an Australian huntsman spider hunting. It's awesome.

[@dibbidydoo4318]

it can't be an accident if it happened multiple times.

[@gasdive]

@@dibbidydoo4318 but 4 legs has only happened once.

[@aceman0000099]

@@gasdive would 8 legs allow for running/sprinting?

[@gasdive]

@@aceman0000099 they seem to. The Australian huntsman spider is *incredibly* fast. They don't spin webs, they chase down their prey.

[@gasdive]

They run at about 5m per second. It's mind blowing to see.

[@TaranovskiAlex]

Sorry, so once again - what do you do specifically? You replicate what boston dynamics did, but from scratch? And why couldn't you use the spot or other models? Why can't you use that chinese model (forgot the name, unfortunately - you can look it up in the news) unveiled not so long ago?

[@tomfoolery2913]

Would you say that there should only be one car company because Henry Ford made the model T?

[@TaranovskiAlex]

@@tomfoolery2913 No, but from the software development perspective - it is reinventing the wheel/bicycle, isn't it? Even for the purpose of research - why not using something that is already there? Unless you have your own idea for one billion dollars...

[@recklessroges]

ANYmal ? MAL ? Really ? Why not just skip to calling it Death Dog 9000 /s

[@johnsenchak1428]

NOT COMPUTER RELATED REPORTED FOR ABUSE