A short YouTube explainer by Dr. Zsolnai (Two Minute Papers) describing a research demonstration where a table-tennis robot learns primarily in simulation (sim-to-real). The robot observes brief human play sequences, builds a simulated opponent model, trains extensively in simulation, transfers the policy to the real robot, and iteratively refines performance by replaying real-world failures back into simulation; results include long rallies (up to 100+ hits) and strong generalization to different human players.
Chapters
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Do you see this new table tennis robot? It barely played any games in the real world, yet it can return the ball more than a hundred times without failing. Wow. So how is this even possible? Well, this is a simtorial paper, which means that first the robot starts learning in a simulation. OpenAI did this earlier by teaching the robot hand in a simulation heated environment to manipulate this ruby cube. And Tesla also trains its cars in a computer simulation. Why? Well, in the real world some things are possible, but in a simulated world anything is possible. Yes, even this. And the self driving car can safely train in this environment. And when it is ready, it can be safely brought into the real world. How cool is that? Now how do we apply this concept to table tennis? Well, in this case the robot would not move, but it would play a computer game in its head, if you will. But not so fast. That is impossible. What are we simulating exactly? The machine doesn't even know how humans play. There is no one to play against. Now check this out. To solve this, first the robot asks for some human data. Look, it won't do anything. It just observes how we play and it only requires these short sequences. Then it builds a model of how we play and embeds us into a computer simulation where it plays against us over and over again without any real physical movement. It is training the brain, if you will. And now comes the key step. This knowledge from the computer simulation is now transferred to the real robot. And now let's see if this computer game knowledge really translates to the real world. So can it return this ball? It can. Well, kind of. One more time. Ok, Better. And now? Well, it missed again. I see some signs of learning here, but this is not great. So is that it? So much for learning in a simulation and bringing this knowledge into the real world, right? Well, do not despair because there is still hope. What can we do? Well, now it knows how it failed and how it interacted with the human. Yes, that is great. Why? Because it can feed this new knowledge back into the simulation. The simulation can now be fired up once again and with all this knowledge it can repeat until the simulation starts looking very similar to the real world. That is where the real fun begins. Why? Well, check this out. This is the previous version of this technique. And as you see, this does not play well. So how about the new method? Now hold on to your papers and marvel at this rally. 82 hits and not one mistake. This is so much better. Wow. This sim 2 reel concept really works. And wait a minute, we are experienced fellow scholars here, so we have a question. If the training set was built from data, when it played against this human being, does it really know how to play against only this person? Or did it obtain more general knowledge? And can it play with others? Well, let's have a look. The robot hasn't played this person before and let's see how the previous technique fares. Well, that was not a long rally and neither is this one. And now let's see the new method. Oh my, this is so much better. It learns much more general information from the very limited human data it was given. So it can play really well with all kinds of players of different skill levels. Here you see a selection of them and all this from learning in a computer game with just a tiny bit of human behavioral data. And it can even perform a rally of over a hundred hits. What a time to be alive. So does this get your mind going? What would you use this sim 2real concept for? Let me know in the comments below.
If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the world for GPU Cloud computer. No commitments or negotiation required. Just sign up and launch an instance and hold on to your papers. Because with Lambda GPU Cloud you can get on demand a 100 instances for $1.10per hour versus $4.10 per hour with AWS. That's 73% savings. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today.
Thanks for watching and for your generous support and I'll see you next time.
