Besides the awkwardness climbing stairs, this is really cool.
That's very fluid and natural movement, I'd imagine this could make game animations alot more believable.
Maybe this could help translate into say Boston Dynamic bipeds be more fluid/natural in walking?
[QUOTE=LoneWolf_Recon;52176662]That's very fluid and natural movement, I'd imagine this could make game animations alot more believable.
Maybe this could help translate into say Boston Dynamic bipeds be more fluid/natural in walking?[/QUOTE]
Boston Dynamics robots seem to move in that unnatural way more because of how they're constructed rather than because of software.
[QUOTE=Broguts;52176720]Boston Dynamics robots seem to move in that unnatural way more because of how they're constructed rather than because of software.[/QUOTE]
I meant more of implementing this neural net in the way for larger scale navigation and integration with the robot's own control loop.
All of BD's bipeds boil down to the same DoF that we as humans have, its just how weight's distributed that's different (Engineers mimic nature and all).
[QUOTE=LoneWolf_Recon;52176734]I meant more of implementing this neural net in the way for larger scale navigation and integration with the robot's own control loop.
All of BD's bipeds boil down to the same DoF that we as humans have, its just how weight's distributed that's different (Engineers mimic nature and all).[/QUOTE]
I doubt it. AFAIK this system works for aesthetic purposes but probably won't be helpful for robots.
[QUOTE=DoctorSalt;52176997]I doubt it. AFAIK this system works for aesthetic purposes but probably won't be helpful for robots.[/QUOTE]
Depends what kind of robots you use it for. Boston Dynamics robots tend to be packmules more than anything else, but robots like [URL="http://www.rethinkrobotics.com/baxter/"]Baxter[/URL] use machine learning and vision (buzzwords yay) to mimic human motion and be less hazardous to work around. We have a Baxter unit at work, its pretty neat to have it mimic your actions but it still moves rather unnaturally - maybe this principle could help make it seem more "natural"?
[QUOTE=LoneWolf_Recon;52176662]That's very fluid and natural movement, I'd imagine this could make game animations alot more believable.
Maybe this could help translate into say Boston Dynamic bipeds be more fluid/natural in walking?[/QUOTE]
Boston Dynamics robots don't need to walk realistically, they only need to walk efficiently.
this has always been one of my biggest annoyances with games, i almost always check to see if my character's feet stay planted at angles. staying planted AND movement being compensated for no matter how the player goes? that's sick. only thing i can see being an issue is it seems to slow down turning speed significantly at times as a result of the interpolation
I'm curious just how many raw animations it requires to be able to generate these final motions.
Nice and all, but I don't really see a semblance of weight transfer when it walks.
[QUOTE=Gmod4ever;52177329]I'm curious just how many raw animations it requires to be able to generate these final motions.[/QUOTE]
Since it's phase based, assumedly not that many.
[QUOTE=LoneWolf_Recon;52176662]That's very fluid and natural movement, I'd imagine this could make game animations alot more believable.
Maybe this could help translate into say Boston Dynamic bipeds be more fluid/natural in walking?[/QUOTE]
Neural networks and reinforcement learning has been employed for a while in robotics. I'd wager Boston Dynamics already uses it to some degree for their robot locomotion projects.
Using a combination of motion-matching and machine learning is a huge leap forward for game animation because once trained, the system can seamlessly transition between animations without having massive blend state graphs for each in-out condition.
[media]https://www.youtube.com/watch?v=VBciHbVP8A4[/media]
Ubisoft is kinda pioneering it right now.
This is a very convincing simulation of a person who has never encountered stairs before.
I'm sure they could have some marker for things like stairs where the next step must land on the next appropriate platform instead of going two steps at a time like a socially awkward nerd tends to do
that's how i usually end up walking up stairs because it's less tiring to skip steps.
fuck your social conventions ^
Not skipping steps is a sign of manlets.
Short legs of short manlets can't efficiently skip steps
[QUOTE=Matrix374;52180214]Not skipping steps is a sign of manlets.
Short legs of short manlets can't efficiently skip steps[/QUOTE]
Tbh I refuse to communicate with people who do not skip at least 1 step when walking up the steps
[editline]3rd May 2017[/editline]
Sorry but umm GO AWAY FREAK
I want to see an animation system like this where the character is interacting with another rigid object, like a rifle. Still self-balancing while having to hold onto it, taking the rifle's weight into account and such. Shooting while scrambling sideways towards cover would look amazing too.
watched it at work so no sound, but i assume this is for the character's movement, foot placement etc?
really neat stuff. i did a course on intelligent systems last semester and have a huge boner for ann (despite the fact that i never got mine to work properly) and things like this makes me want to get back into it.
really cool stuff about the phase function. what we were taught was you can have a simple linear function or euclidean function, but they only really taught us only one function and asserted that thats what you'll need for a neural network.
which is great because this method is actually close to a concept i had for bots in a multiplayer game that learns from people playing it. i figured you'd only get so far with decision based ai and if you designed the neural network right, you'd be able to train it with replays from say a game of cs at different skill levels and then it would be able to mimic the actions players accurately to be convincing enough to appear to be a player of that skill level.
[QUOTE=Pandamobile;52177363]Neural networks and reinforcement learning has been employed for a while in robotics. I'd wager Boston Dynamics already uses it to some degree for their robot locomotion projects.
Using a combination of motion-matching and machine learning is a huge leap forward for game animation because once trained, the system can seamlessly transition between animations without having massive blend state graphs for each in-out condition.
[media]https://www.youtube.com/watch?v=VBciHbVP8A4[/media]
Ubisoft is kinda pioneering it right now.[/QUOTE]
The guy also did a GDC talk on this
[media]https://www.youtube.com/watch?v=KSTn3ePDt50&ab_channel=KristjanZadziuk[/media]
[QUOTE=Pat.Lithium;52180373]
really neat stuff. i did a course on intelligent systems last semester and have a huge boner for ann (despite the fact that i never got mine to work properly) and things like this makes me want to get back into it.
[/QUOTE]
Out of curiosity, what was the goal? Did you implement backprop, or other forms of activation/error functions, or use some frameworks to solve something?
Sorry, you need to Log In to post a reply to this thread.
