Main video accompanying the SIGGRAPH 2017 paper: "DeepLoco: Dynamic Locomotion Skills Using Hierarchical Deep Reinforcement Learning". Project page: www.cs.ubc.ca/~van/papers/2017...
Пікірлер: 32
@Frautcres7 жыл бұрын
2:42 Didn't see that coming.
@Waffle45697 жыл бұрын
*wasted*
@danthemango6 жыл бұрын
*pow* right in the kisser
@ForgetfulHatter7 жыл бұрын
2:42 Skynet will remember that.
@azraelle62327 жыл бұрын
I wonder what would happen if you had two figures with opposing goals for the ball? One trying to get it to the red goal, the other trying to get it to a green goal.
@phlimy7 жыл бұрын
I agree! I would love to see the result.
@Napert5 жыл бұрын
They would probably just push each other till the end of time
@Skyliner_3695 жыл бұрын
If possible, I'd recommend training with arms and a head to help with balance as well as joint force instead of actual angle maps. Heck, train WITH rough terrain.
@behrTheNerd7 жыл бұрын
So, he's left footed, short and has quirky geometry, dodges obstacles on a field and gets attacked. Just put a Barca jersey on it, Mike. We all know you're making Lio Messi. Maybe soccer nets as a proper reward function for the poor feller?
@MrVersion217 жыл бұрын
Hi Michiel, I just stumbled onto your work here on KZfaq. I think its inspiring! I am working with humanoid robots in my research and one of the main challenges we face is that the robots break a lot. So getting a lot of training data is not easy. What parts do you think are needed to transfer your work on a real robot? Can the training be started in simulation and then transferred to the robot?
@m.vandepanne7 жыл бұрын
Indeed, enabling the transfer of control policies from simulation to robots in the real world is a problem that is of interest to many right now. Possible solutions include: (a) learning policies in simulation which are valid for an ensemble of model parameters, i.e., are robust to some expected variation; (b) moving away from model-free learning methods towards more model-based learning methods; (c) developing "safe" learning methods; (d) learning better forward dynamics by predicting the difference between a baseline simulation and the actual observed dynamics; (e) learning on smaller, more-robust robots; and no doubt many other ideas that I have not even touched on. Many groups have already demonstrated tangible progress in (deep) learning on real robots, most often with approaches that leverage model-based learning methods. In the meantime, the learned control policies provide an indication of what current simulation-based models are capable of, how natural the motions can be, and the size of learned policies. And these simulated models can then already be used in simulations, games, visualizations, etc. Overall, the field is moving very quickly, so I think that we'll see many advances in the coming months and years!
@Fuglyuck7 жыл бұрын
have you though about adding the top half of the body to allow for more complex tasks that require more balance? Such as steeper inclines etc etc
@m.vandepanne7 жыл бұрын
yes, more skills in moving through more complex environments, using the hands, body, and knees as required, is an exciting (and difficult!) direction for future work...
@dancre7 жыл бұрын
And how about trying some stairs or some jumping
@ValentinHarbinger7 жыл бұрын
Can you tell me what minimum qualifications and background (apart from a Master's degree) do you expect from a future PhD student in this area? I've been fascinated by this kind of work for more than 6 years since coming across the work of Reil and Sims. Seriously, what can I do until December that would boost my chances to get into this research group? There is nothing that I would not do.
@m.vandepanne7 жыл бұрын
In this area, it is helpful to have knowledge of animation, machine learning, physics-based simulation, robotics, and even some biomechanics and motor control. No-one comes into this area with all this knowledge. So perhaps what is more important is to have taken some mix of relevant undergrad courses, and to have a demonstrated ability to lead-and-execute on projects of your own creation.
@ericbochat6 жыл бұрын
Hi, i'm a 19 yo student (+2 years post graduation) in the following fields : Mathematics / Physics / Engineering science, it indeed sounds really vague, but i can't get more precise since the 2 years i'm in are a global preparation to Engineering Schools of the highest reputation here in France. We only specialize our selfs in specific fields there. If I understood your comment well, what I should do now is keep doing my best in order to join the best Engineering School possible right? Anyway, the thought of working with you guys, or in a similar field motivates me a lot. Motion and Animation always interested me and I always wanted to be more than a software user. Thanks for your amazing work !
@lavenderglaab8356 жыл бұрын
2:33 Now that's comedy
@adamjanuszewski38427 жыл бұрын
maybe you can add hand finger movements, or just mitten like hands... to grab certain objects and move them to a certain location for example: the character needs to grab a cylinder somewhat like a soda can and move it from one table, to the other. cool idea right? and we humans also balance with our arms like when we are about to fall over a steep cliff and we pull our arms back to regain balance back to the flat surface. maybe you can use that for difficult tasks, like in climbing where you have to grab and step cliff hangers on a steep cliff to climb up, which you need both arms and legs for it. phew... that was alot well hope its not too much work for you... you can just do the grabbing simulation part... not the climbing... dont want to stress you out... xD well hope you like the idea!!!
@m.vandepanne7 жыл бұрын
Indeed, locomotion is just the start -- many more skills are needed! And then they need to be integrated. The months & years ahead will be interesting!
@adamjanuszewski38427 жыл бұрын
its interesting when you think about a town building game with locomotive characters like in the video trying to use an axe and pick up logs with physics, if u imagine, the enviroment, trees, rocks, mountains, and imagine the ragdolls walking picking up logs that has weight. and just transporting them in the warehouse! that could be awesome, heck! it could be a new top rated game for best civilization physics. just an idea tho. probably for someone else to make. but who knows you can do it to!
@knowlen6 жыл бұрын
"Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations" (Rajeswaran, Kumar et al., 2017) /watch?v=jJtBll8l_OM
@endingalaporte5 жыл бұрын
Shit gets Real
@repogamesstudio23665 жыл бұрын
cool
@Dataism6 жыл бұрын
I wonder if Pixar uses this tech.
@RaymondLei_yunyunzai6 жыл бұрын
4:32 RIP...
@ayachavez74505 жыл бұрын
Feels like low qual sims
@pokey20396 жыл бұрын
2:42 When puberty hits you hard :P
@martinenriquedomingueznarc97365 жыл бұрын
2:43 poor guy haha
@paulgabel82617 жыл бұрын
Now put Michael Jackson Thriller and watch this video