It isn't enough to have artificial intelligence tanning humanity's hide at every board game in existence, it has to be done at ping pong as well. At the rate these things improve, it will take on pros in no time.
The i-Sim2Real project is about building a robotic system that can work with and around fast paced and unpredictable human behavior. A balance of complexity and simplicity is what table tennis has in it's favor.
An artificial intelligence creation process in which a machine learning model is taught what to do in a virtual environment and then applies that knowledge in the real world is called "Sim2Real". When it takes years of trial and error to arrive at a working model, it is necessary to do it in a sim so that years of real-time training can occur in a few minutes or hours.
It isn't always possible to do something in a sim, for example if a robot needs to interact with a human. Real-world data is needed to start with. You don't have the human data because you need it to make the robot the human would interact with and generate that data
This pitfall was avoided by the researchers by starting simple.
[i-Sim2Real] uses a simple model of human behavior as an approximate starting point and alternates between training in simulation and deploying in the real world. In each iteration, both the human behavior model and the policy are refined.
The robot is only just beginning to learn so it is okay to start with a bad approximation. Every game collects more real human data to improve accuracy and learn more.
The team's table tennis robot was able to carry out a large rally because of the approach. You can check it out.
It was able to return the ball to different regions because it was good enough to start executing a strategy.
The team tried to return the ball to a specific spot from a variety of positions. This isn't about creating the ultimate ping pong machine, but finding ways to efficiently train with and for humans without making them repeat the same action thousands of times.
In the summary video, you can learn more about the techniques the team used.
All Rights Reserved © 2024
