After playing more than 300 times in four minutes, Google taught its robot to play table tennis

Let a table tennis enthusiast play against a robot. Judging from the development trend of robots, it’s really hard to say who will win and who will lose. ​

The robot has dexterous maneuverability, flexible leg movement, and excellent grasping ability... and has been widely used in various challenging tasks. But how do robots perform in tasks that involve close interaction with humans? Take table tennis as an example. This requires a high degree of cooperation from both parties, and the ball moves very fast, which poses a major challenge to the algorithm.

In table tennis, the first priority is speed and accuracy, which places high demands on learning algorithms. At the same time, this sport has two major characteristics: highly structured (with a fixed, predictable environment) and multi-agent collaboration (robots can fight with humans or other robots), making it an ideal place to study human-computer interaction and reinforcement learning. An ideal experimental platform for problems.

The robotics research team from Google has built such a platform to study the problems faced by robots learning in multi-person, dynamic and interactive environments. Google also wrote a special blog for this purpose to introduce the two projects they have been studying, Iterative-Sim2Real (i-S2R) and GoalsEye. i-S2R enabled the bot to play over 300 matches with human players, while GoalsEye enabled the bot to learn some useful strategies (goal-conditional strategies) from amateurs.

i-S2R strategy allows robots to compete with humans. Although the robot’s grip does not look very professional, it will not miss a ball:

You come and I go, it’s pretty much the same thing, it feels like I’m playing a high-quality ball.

The GoalsEye strategy can return the ball to the designated position on the table, just like hitting it wherever you point it:

i-S2R: Playing games with humans using simulators

In this project, the robot aims to learn to cooperate with humans, that is, to play sparring with humans for as long as possible. Since training directly against human players is tedious and time-consuming, Google adopted a simulation-based approach. However, this faces a new problem. It is difficult for simulation-based methods to accurately simulate human behavior, closed-loop interaction tasks, etc.

In i-S2R, Google proposed a model that can learn human behavior in human-computer interaction tasks and instantiated it on a robotic table tennis platform. Google has built a system that can achieve up to 340 batting throws with amateur human players (shown below).

Human and robot fight for 4 minutes, up to 340 times back and forth

Learning Human behavior models

Allowing robots to accurately learn human behavior also faces the following problems: If there is not a good enough robot strategy from the beginning, it is impossible to collect high-quality data about how humans interact with robots. But without a human behavior model, the robot strategy cannot be obtained from the beginning. This problem is a bit convoluted, like which came first, the chicken or the egg. One approach is to train robot policies directly in the real world, but this is often slow, costly, and poses safety-related challenges that are further exacerbated when humans are involved.

As shown in the figure below, i-S2R uses a simple human behavior model as an approximate starting point and alternates between simulation training and real-world deployment. In each iteration, human behavior models and strategies are adapted.

i-S2R Method

Google breaks down the results of the experiment based on player type: Beginners (40% of players), Intermediate (40% of players) and Advanced (20% of players). From the experimental results, i-S2R performs significantly better than S2R FT (sim-to-real plus fine-tuning) for both beginners and intermediate players (80% of players).

i-S2R results by player type​

GoalsEye: Hit exactly where you want

In GoalsEye, Google also demonstrated a method that combines behavioral cloning techniques to learn precise targeting strategies. . ​

Here Google focuses on the accuracy of table tennis. They hope that the robot can accurately return the ball to any designated position on the table, as shown in the figure below. To achieve the following effects, they also used LFP (Learning from Play) and GCSL (Goal-Conditioned Supervised Learning).

The GoalsEye strategy targets a 20cm diameter circle (left). A human player could aim for the same target (right) ​

In the first 2,480 demos, Google’s training strategy only worked 9% of the time Accurately hit a circular target with a radius of 30 cm. After about 13,500 demonstrations, the accuracy of the ball reaching its target increased to 43 percent (bottom right).

​

For more introduction to these two projects, please refer to the following link:

• Iterative-Sim2Real Home Page: https://sites.google.com/view/is2r
• ##GoalsEye Home Page: https://sites.google.com /view/goals-eye

The above is the detailed content of After playing more than 300 times in four minutes, Google taught its robot to play table tennis. For more information, please follow other related articles on the PHP Chinese website!