i like the part where he steps over the obstacle
May 2, 2017 6:20 PM Subscribe
Phase-Functioned Neural Networks for Character Control, aka holy shit look at this character animation. If you want some more details, check out the research paper by Daniel Holden, Taku Komura, and Jun Saito.
Holy shit.
SIGGRAPH is one of those conferences I can't imagine I'll ever submit to, but would love to attend once in my life.
posted by codacorolla at 6:37 PM on May 2, 2017 [4 favorites]
SIGGRAPH is one of those conferences I can't imagine I'll ever submit to, but would love to attend once in my life.
posted by codacorolla at 6:37 PM on May 2, 2017 [4 favorites]
Looks like Daniel finally turned his life around - oh, not that Daniel Holden.
But seriously, this seems pretty cool. If someone can summarize how the phase function works in layman's terms, I'd be mighty appreciative (looking at the paper now)
posted by piyushnz at 6:59 PM on May 2, 2017
But seriously, this seems pretty cool. If someone can summarize how the phase function works in layman's terms, I'd be mighty appreciative (looking at the paper now)
posted by piyushnz at 6:59 PM on May 2, 2017
The paper looks really interesting and I'm impressed with how they got there, but the results in the demo video look... kind of not great to me? (I may have the wrong eyes for this, since I just came off of playing Kentucky Route Zero, where the way a character runs, skips, or trudges is basically a narrative device.)
posted by phooky at 7:15 PM on May 2, 2017
posted by phooky at 7:15 PM on May 2, 2017
Watching that guy bumble about that alien terrain made me anxious he'd never get around to riding to Lexington to warn them that the British were coming.
posted by ejs at 7:25 PM on May 2, 2017 [29 favorites]
posted by ejs at 7:25 PM on May 2, 2017 [29 favorites]
To me he looked like he was right on the edge of being TOTALLY DONE with chasing around that stupid arrowsnake all over freakin' creation.
posted by Greg_Ace at 7:30 PM on May 2, 2017 [3 favorites]
posted by Greg_Ace at 7:30 PM on May 2, 2017 [3 favorites]
I agree that the demo doesn't look particularly realistic, but I think it's partially an issue with using a clothed figure with no cloth physics. It's difficult to separate any evaluation of the figure's movement from the stiff, unrealistic clothing. They should have used an androgynous, unclothed figure to avoid such distractions.
posted by xyzzy at 7:32 PM on May 2, 2017 [2 favorites]
posted by xyzzy at 7:32 PM on May 2, 2017 [2 favorites]
using a clothed figure with no cloth physics
Wow, that's definitely a factor. I really do wonder why they chose that model.
posted by phooky at 7:35 PM on May 2, 2017
Wow, that's definitely a factor. I really do wonder why they chose that model.
posted by phooky at 7:35 PM on May 2, 2017
*squints*
Is the yellow dude supposed to be Mayor Hancock or have I been playing too much Fallout?
posted by a power-tie-wearing she-capitalist at 8:10 PM on May 2, 2017 [2 favorites]
Is the yellow dude supposed to be Mayor Hancock or have I been playing too much Fallout?
posted by a power-tie-wearing she-capitalist at 8:10 PM on May 2, 2017 [2 favorites]
I understand most of the machine learning theory explained here but holy fuck.
posted by Young Kullervo at 8:21 PM on May 2, 2017 [1 favorite]
posted by Young Kullervo at 8:21 PM on May 2, 2017 [1 favorite]
One step closer to my dream of being able to make a big Hollywood production, by myself, in my pajamas.
posted by bigbigdog at 8:23 PM on May 2, 2017 [3 favorites]
posted by bigbigdog at 8:23 PM on May 2, 2017 [3 favorites]
If someone can summarize how the phase function works in layman's terms, I'd be mighty appreciative (looking at the paper now)
Ok, so, a little bit of animation theory. There's four key portions of a walk or jog: right leg up, right leg down, left leg up, and left leg down. How you plan for and react to an upcoming obstacle or change of direction depends heavily on what part of that cycle you're in. It's a kind of rule-based constriction: if your right leg is in the air, you have to put it down before starting to lean right.
Neural networks are typically very bad at rules-based constrictions: they know about patterns and gradients, not so much about limits and phase transitions. See the Standard Neural Network portion of the video on their webpage, where it clearly doesn't understand that a foot on the ground needs to stay at that point on the ground. It just tries to move closer to the ideal state, by whatever line is quickest, without a clear way of forcing a repeated order onto the movement.
What they're doing with the phase function is smoothly transitioning between four different rules of motion, one for each of the four quadrants of the walk. They do this by essentially training four different neural networks, then gradually moving control between them in determining what the next movement is. This way, the right-leg-up network will never try to slide the left leg along the plane, instead waiting for the left-leg-up network to cycle into control and solve that problem. It's a way to impose an extra constraint about the world onto a tool that usually goes wonky when trying to deal with them.
posted by persona at 8:38 PM on May 2, 2017 [48 favorites]
Ok, so, a little bit of animation theory. There's four key portions of a walk or jog: right leg up, right leg down, left leg up, and left leg down. How you plan for and react to an upcoming obstacle or change of direction depends heavily on what part of that cycle you're in. It's a kind of rule-based constriction: if your right leg is in the air, you have to put it down before starting to lean right.
Neural networks are typically very bad at rules-based constrictions: they know about patterns and gradients, not so much about limits and phase transitions. See the Standard Neural Network portion of the video on their webpage, where it clearly doesn't understand that a foot on the ground needs to stay at that point on the ground. It just tries to move closer to the ideal state, by whatever line is quickest, without a clear way of forcing a repeated order onto the movement.
What they're doing with the phase function is smoothly transitioning between four different rules of motion, one for each of the four quadrants of the walk. They do this by essentially training four different neural networks, then gradually moving control between them in determining what the next movement is. This way, the right-leg-up network will never try to slide the left leg along the plane, instead waiting for the left-leg-up network to cycle into control and solve that problem. It's a way to impose an extra constraint about the world onto a tool that usually goes wonky when trying to deal with them.
posted by persona at 8:38 PM on May 2, 2017 [48 favorites]
Does the character have to wear a hat for this animation to work properly, or no?
posted by stargell at 9:04 PM on May 2, 2017 [2 favorites]
posted by stargell at 9:04 PM on May 2, 2017 [2 favorites]
Karlgren proved in the 1980s that the result holds for any regular n-corn.
posted by No-sword at 9:10 PM on May 2, 2017 [14 favorites]
posted by No-sword at 9:10 PM on May 2, 2017 [14 favorites]
codacorolla: SIGGRAPH is one of those conferences I can't imagine I'll ever submit to, but would love to attend once in my life.
I attended back in the late 1990s, but as an idiot kid in high school who had an awesome opportunity to play around at Alias|Wavefront, but didn't realize what that meant. We had oold systems and were playing around with modeling and animation, but the most I made was a weird kangaroo-mouse looking thing that might have moved, somewhat. The best in our group was a guy who made a bicycle chain animation, but to render it, the folks at A|W ran it on their lower-class render farm after hours, because our computers were pretty much crap (what else would you give random high school students?). Anyway, we went down to SIGGRAPH with the idea of finding a ton of mousepads. Yeah, we were idiot kids. Anyway, we found one in the entire venue, but we also picked up some product demo CDs and demo presentation VHS tapes, and some books that we didn't really understand, but hey - they were free!
Youth is wasted on the young.
phooky: I really do wonder why they chose that model.
stargell: Does the character have to wear a hat for this animation to work properly, or no?
I thought they chose the dude-type character in the tricorner hat because you could tell which way they were going.
posted by filthy light thief at 9:24 PM on May 2, 2017 [2 favorites]
I attended back in the late 1990s, but as an idiot kid in high school who had an awesome opportunity to play around at Alias|Wavefront, but didn't realize what that meant. We had oold systems and were playing around with modeling and animation, but the most I made was a weird kangaroo-mouse looking thing that might have moved, somewhat. The best in our group was a guy who made a bicycle chain animation, but to render it, the folks at A|W ran it on their lower-class render farm after hours, because our computers were pretty much crap (what else would you give random high school students?). Anyway, we went down to SIGGRAPH with the idea of finding a ton of mousepads. Yeah, we were idiot kids. Anyway, we found one in the entire venue, but we also picked up some product demo CDs and demo presentation VHS tapes, and some books that we didn't really understand, but hey - they were free!
Youth is wasted on the young.
phooky: I really do wonder why they chose that model.
stargell: Does the character have to wear a hat for this animation to work properly, or no?
I thought they chose the dude-type character in the tricorner hat because you could tell which way they were going.
posted by filthy light thief at 9:24 PM on May 2, 2017 [2 favorites]
Automatically preventing characters from accidentally stepping off into gaping chasms? It'll never catch on.
posted by ckape at 10:15 PM on May 2, 2017 [2 favorites]
posted by ckape at 10:15 PM on May 2, 2017 [2 favorites]
Is the yellow dude supposed to be Mayor Hancock or have I been playing too much Fallout?
I can totally see someone getting so fed up with the walk animations in Fallout 4 that they create something like this.
posted by um at 11:45 PM on May 2, 2017
I can totally see someone getting so fed up with the walk animations in Fallout 4 that they create something like this.
posted by um at 11:45 PM on May 2, 2017
Assuming there's folks here who understand how all this works: how difficult would it be to implement this algorithm in existing game engines? And would this be applicable to walking robots, or are the algorithms completely different there?
wondering whether to expect a tricorn in a future Boston Dynamics video
posted by vanar sena at 12:58 AM on May 3, 2017
wondering whether to expect a tricorn in a future Boston Dynamics video
posted by vanar sena at 12:58 AM on May 3, 2017
I imagine it would be a middle ware product like the euphoria engine used in gta and sports games.
posted by empath at 3:51 AM on May 3, 2017 [1 favorite]
posted by empath at 3:51 AM on May 3, 2017 [1 favorite]
The narrator sounds like Brett Domino.
posted by grumpybear69 at 5:28 AM on May 3, 2017 [2 favorites]
posted by grumpybear69 at 5:28 AM on May 3, 2017 [2 favorites]
AI - this idea that computers will "outthink" and replace 50% of the workforce in under a decade is instead going to have FAR more cycles instead spent on providing entertainment rather than provide the work reducing/removal of drudgery that was promised in the 1920's when mass production would create more leisure time?
Or is the knock-off effect of providing entertainment in magical AI graphics cards going to instead get re purposed as the outthink/50% reduction devices the same way graphic cores are used for vector math/mining bitcoins/creating password hashes. Or how the data storage needs of the surveillance state is making hard drives cheap and big?
posted by rough ashlar at 6:33 AM on May 3, 2017
Or is the knock-off effect of providing entertainment in magical AI graphics cards going to instead get re purposed as the outthink/50% reduction devices the same way graphic cores are used for vector math/mining bitcoins/creating password hashes. Or how the data storage needs of the surveillance state is making hard drives cheap and big?
posted by rough ashlar at 6:33 AM on May 3, 2017
The annual SIGGRAPH melting of minds.
posted by GuyZero at 7:42 AM on May 3, 2017 [1 favorite]
posted by GuyZero at 7:42 AM on May 3, 2017 [1 favorite]
I don't know, rough ashlar, but yesterday a physiotherapist told me about talking to a roboticist at a physio convention who is working on robots with perfect-human movement because they think being around such robots would make a lot of humans healthier.
And then we talked about posture and carriage and gait that we learn from people around us, and I did an 1830s dance step and a 1930s dance step, and she showed me how English and French and German people stand differently when waiting in line.
posted by clew at 12:17 PM on May 3, 2017 [1 favorite]
And then we talked about posture and carriage and gait that we learn from people around us, and I did an 1830s dance step and a 1930s dance step, and she showed me how English and French and German people stand differently when waiting in line.
posted by clew at 12:17 PM on May 3, 2017 [1 favorite]
Looks like the lead author (Daniel Holden) works at Ubisoft Montreal. His twitter looks very interesting.
posted by tickingclock at 3:35 PM on May 3, 2017 [1 favorite]
posted by tickingclock at 3:35 PM on May 3, 2017 [1 favorite]
Clew - if you "learn from people around" robots may not be the lowest cost way. Strapping on VR and watching the photo-realistic images would seem to be cheaper.
posted by rough ashlar at 4:56 PM on May 3, 2017
posted by rough ashlar at 4:56 PM on May 3, 2017
I'm not going to test that, rough ashlar, but I expect someone will. Do immersed gamers imitate their avatars now?
posted by clew at 10:25 PM on May 3, 2017
posted by clew at 10:25 PM on May 3, 2017
clew, that sounds like a really fun physiotherapist session. I would like to know which nationality I stand the most like!
The little guy is a far better hiker than I am. Rock fields usually mean I am on all fours, and I think the guy only stabilizes himself with a hand once.
posted by batter_my_heart at 10:26 PM on May 3, 2017
The little guy is a far better hiker than I am. Rock fields usually mean I am on all fours, and I think the guy only stabilizes himself with a hand once.
posted by batter_my_heart at 10:26 PM on May 3, 2017
oh yay is it already siggraph paper time?! One of my favourite times of the year :)
posted by danhon at 10:56 PM on May 3, 2017 [1 favorite]
posted by danhon at 10:56 PM on May 3, 2017 [1 favorite]
« Older This Is Not Your Grandma's Knitting | Draping dynamic contours with a drooping material... Newer »
This thread has been archived and is closed to new comments
posted by Samizdata at 6:35 PM on May 2, 2017