Yesterday, Google announced that it is working on a "ping-pong robot" project that can catch the ball 340 times in one round when playing against humans.
At present, Google emphasizes that this is just a "cooperation" between humans and AI, not to defeat humans, but with the speed of AI growth, it will soon start against professional players.
The project, called i-Sim2Real, isn't just about ping pong, it's about building a robotic system that can work with fast-paced and relatively unpredictable human behavior. The advantages of table tennis are that it is fairly restricted (compared to playing basketball or cricket) and strikes a balance between complexity and simplicity. In the process, machine-learning models are taught what to do in a virtual environment or simulation and then apply that knowledge, with the goal of playing catch rounds with humans for as long as possible without making a mistake.
I addition to the i-Sim2Real method of alternating simulation and reality, Google researchers are also exploring a method of learning using only real data, namely the GoalsEye project, the former can alternate learning strategies in simulation and reality, While the latter learning from real-world unstructured data, combined with self-training, is effective for learning target-conditioned policies with precise and dynamic requirements.
0 Comments