ping pong paddle and ball

How Google taught bots to play ping pong

yesterday, Google Research revealed two new projects He was working on a robot that plays table tennis. The Google Robotics Team I learned a robotic arm to play over 300 rally rallies with other people and accurately return the “human hobbyist”. While this may not sound impressive given how poorly some people perform in table tennis, the same methods can be used to train robots to perform “high-speed, dynamic tasks” that require close human-robot interaction.

Table tennis is an interesting task for robots to learn because of two complementary properties: It requires quick and precise movements in a structured game that takes place in a stable and predictable environment. The learning algorithm that a bot relies on to make decisions has to work hard to be good, but the limits of table tennis limit how much of the world it has to deal with. It helps that playing table tennis is a two-part task: a robot can play with another robot (or a simulator) or a human can train it. All this makes it a great setting for exploring human-robot interaction reinforcement learning techniques (Where the robot learns from the action).

Google engineers designed two separate projects using the same robot. Recursive Sim2Realwhich is going to be Presented in CoRL Later this year, the Objectiveswhich is going to be Filed in IROS next week. Iterative-Sim2Real is the software that has trained the robot to play 300 co-operative rounds with humans while GoalsEye allows retransmission to a specific target point on the table with hobby-like precision.

Sim2Real iterations is an attempt to get around the “chicken and egg problem” of teaching machines to imitate human behaviours. The research team explains that if you don’t have a good bot policy (a bot set of rules) to start with, you can’t collect high-quality data about how people interact with it. But without a human behavior model to begin with, you can’t come up with a bot policy in the first place. One alternative solution is to train the bots exclusively in the real world. However, this process is “often slow, costly, and poses safety challenges, which are further exacerbated when people are involved.” In other words, it takes a long time and people can get hurt by swinging the robot’s arms with table tennis rackets around.

Iterative-Sim2Real avoids this problem by using a very simple model of human behavior as a starting point and then training the bot on both the simulation and the human in the real world. After each iteration, both the human behavior model and the bot policy are revised. Using five human subjects, the robot trained with Iterative-Sim2Real outperformed an alternative method called Sim to real plus refine. He had fewer walks that ended with less than five shots and had an average run length of 9 percent longer.

On the other hand, GatesEye addressed a different set of training problems and taught the robot to return the ball to a random location such as “back left corner” or “above the net on the right side.” Learning by imitation – where a robot develops a game strategy that is derived from human performance data – is difficult to implement in high-speed settings. There are so many variables that affect how a human hits a ping-pong ball that keeping track of everything necessary for the robot to learn is practically impossible. Reinforcement learning is usually fine for these situations but can be slow and ineffective – especially in the beginning. (In other words, it takes a lot of repetition to develop a somewhat limited gameplay strategy.)

GoalsEye attempts to overcome both sets of problems with a “small, poorly organized, untargeted” raw dataset that enables the bot to learn the basics of what happens when it hits a ping-pong ball and then let it practice on itself. To teach her to hit the ball accurately to specific points. After being trained in 2,480 initial runs, the robot was able to return the ball to a distance of 30 cm (roughly one foot) in just 9 percent of the time. But after self-training for roughly 13,500 rounds, it was accurate 43 percent of the time.

While teaching bots to play games may seem trivial, the research team maintains that solving these types of training problems with table tennis has potential real-world applications. Iterative-Sim2Real allows bots to learn from interactions with humans while GoalsEye demonstrates how bots can learn from unstructured data and practice autonomously in a ‘micro, dynamic setting’. Worst case scenario: If Google’s big goals aren’t met, they can at least build a robotic table tennis coach.


#Google #taught #bots #play #ping #pong

Leave a Comment

Your email address will not be published. Required fields are marked *