now that Reinforcement Learning is in Unity, does that mean that dynamic, physics-based locomotion will be possible?
http://www.cs.ubc.ca/~van/papers/2017-TOG-deepLoco/
…reminds me of certain someone who really loves talking about double precision.
You need to consider overhead of any deep learning based system. Deep learning and neural nets in essence mean a LOT of floating point operations per frame. This will limit number of agents you might have in the level.
Then there’s dealing with sample/training data.
1 Like
Physics based locomotion was already possible.
3 Likes
Physics based animation is definitely the future of animation, since it is one of the most difficult, time-consuming and costly parts of making a game.
However it’s important to consider that what is really useful is not the AI/learning algorithm itself, but the fixed result that it produces. Any sort of runtime learning, outside of a few sorts of games, is not just a barrier to the player’s immediate enjoyment (nobody wants to sit around and wait while the AI learns how to be fun) but also an opportunity for a lot of things to go wrong that would be very difficult to debug. Therefore the question is how to produce something flexible and ready with this method, that could be dropped into any game and do a good job of locomotion immediately and consistently.
If full physics-based animation is prohibitive for any reason, there’s still a huge opportunity to improve ‘active ragdoll’ states, such as with a sort of nervous system that contains physics-based reflexes and/or damage states for various situations that blend with animations to make them more reactive to the world.
Anyway, some kind of physics based animation was on Unity’s research roadmap, but I don’t see it there any more.
1 Like
On a scale from 1 to 10 … how applicable to a real physical robot is this learning system?
Are arduino walking robots going to be a thing? I’ve been wanting to play with arduino (for really simple things) for some time now drools
On scale 1 to 10 it is zero.
Game engines have access to underlying structure of the world, while robots deal with unreliable and noisy data.
Arduino itself may be a bit underpowered for a walking robot - due to limited clock speed and amount of memory. However it could work well in conjunction with raspberry pi.
It also may be possible to create a locomotion cycle for arduino robot using genetic algorithm, but this will not create a deep-learning-based walking.
Basically, trying to make an arduino walker most likely will result in you blowing several hundrds of USD on parts.
1 Like
I dont think this will be a thing, especially considering how Arduino is for small things. You have to realise that walking is a very difficult beast- And in a Game as one guy already said the system already knows the whole world through and through. In real life it has to view the world based on motion sensors, cameras and audio, something a machine finds very difficult to do itself. You would need a Computer that is the size of the Japanese K- And its a device that takes up entire rooms of space to even walk in a stable manner. Our Brains do this somehow by firing Neural Oscillations (Brain Waves) in a complex pattern that no one knows much about yet, but on the bright side, Neuromorphic Chips (Neural Networks built with electrical components rather than coded software) with Neuron like circuitry may come out soon and if somehow was able to work with open source hardware (Like Arduino) your dream would be one step closer to becoming a reality 
1 Like
One day I want to do that diy drone as a hobby
1 Like