2024/08/09 13:20
Table tennis AI robot wins over humans! Flexible conversion of forehand and backhand, can catch high balls in tennis, professional coaching: reach the level of intermediate players
2024/08/09 13:20
The table tennis team competition at the Paris Olympics is in full swing, and Google Robot has applied to compete-
The first robot agent to reach the level of human competition has been released!
You see that if you don't pay attention, you win a ball by a professional coach!
The forehand and backhand are quickly converted, and the continuous attack is not a problem~
In the face of some sudden tactics, such as long balls and high balls to wipe tennis, they can also deal with it calmly.
In the actual test, the robot adapted to different player styles in real time, and finally won all the matches against beginners, and also had a 55% win rate against intermediate players.
Little ping-pong ball, take it!
Barney J. Reed, the United States table tennis star against him, spoke highly of it: beyond expectations, the robot has reached an intermediate level.
After watching its performance, netizens said: Can you buy it? Want.
Table tennis is a sport that requires a lot of physical strength, strategy, skills and other aspects, and humans often need to master it after years of training.
Therefore, unlike pure strategy games such as chess and Go, table tennis has become an important benchmark for robots to test their comprehensive capabilities, such as high-speed movement, real-time precise control, strategic decision-making, system design, and so on.
For example, in the face of different landing points of the ball, the robot needs to move its position quickly; In the face of an obvious out-of-bounds ball, the robot should choose not to catch it.
The team found 29 table tennis players of different skill levels to compete, including beginner, intermediate, advanced and above.
Humans played 3 matches against robots, and the game followed standard table tennis rules. (However, since the robot cannot serve, the whole game is served by humans).
Prior to this, there was actually a corresponding research on table tennis robots, and the special thing about this Google robot is that it can have a full-scale competitive duel with humans who have never seen it before.
It can adapt quickly to all kinds of human playing styles.
For example, looking at this player, at the beginning of the game, the robot was obviously still in the process of adapting, and the human defeated the robot with a score of 9 to 2.
But just after the next game, the robot is clearly familiar with the opponent's style and is always chasing the score. The two sides also fought back and forth.
In the end, out of all opponents, the robot won all beginner competitions, with a 55% win rate against intermediate players.
Although there is no way for the robot to defeat the advanced players at the moment, it can be seen from the various feedback from humans that everyone is happy to play with this robot.
Before introducing the method, let's take a look at the hardware configuration of the table tennis robot.
The main body uses a 6-degree-of-freedom ABB 1100 robotic arm from the Switzerland company, mounted on two Festo linear guides, which enable it to move in a plane. The transverse moving guide is 4 meters long, and the longitudinal moving guide is 2 meters long.
The robotic arm is equipped with a 3D-printed racket handle and a racket covered with short-grain rubber.
It's such a little Dency, how did he learn to play table tennis?
In summary, a blended training method combining reinforcement learning and imitation learning was used.
The team designed a hierarchical and modular policy architecture, with the Agent including a low-level skill library (LLC) and a high-level controller (HLC).
LLC is a specialized set of strategies, each of which is trained to perform specific table tennis skills such as forehand shots, backhand shots, serves, etc. These LLCs use a CNN architecture and are trained on an evolutionary strategy algorithm in a simulation environment.
Ball state datasets collected from the real world are used in the training process to ensure consistency between the simulated environment and the real environment.
And the HLC is responsible for selecting the most suitable LLC for each incoming ball.
There are several components: a style strategy to choose a forehand or backhand; a spin classifier to identify the type of spin of an incoming ball; LLC skill descriptors, which describe the capabilities of each LLC; A set of heuristic strategies for shortlisting candidate LLCs based on the current situation.
HLC also uses the LLC preference for online learning to adapt to the characteristics of the adversary and bridge the gap between simulation and reality.
Specifically, the team first collected a small amount of human game data, set the initial task conditions, and then used reinforcement learning to train an agent in a simulated environment, and then deployed the strategy to the real world with zero shots.
The MuJoCo physics engine is used to accurately simulate ball and robot dynamics, including air resistance, Magnus effect, etc., and the topspin ball "correction" is also designed to simulate the real-world topspin and downspin effects by switching different racket parameters in the simulation.
In the process of continuous confrontation between the agent and humans, more training task conditions can be generated, and the training and deployment can be repeated.
Robot skills gradually improve, and the competition gradually becomes more complex, but still based on real-world mission conditions. Once the robot has collected data, it can also identify deficiencies in its capabilities and then compensate for these deficiencies by continuously training in a simulated environment.
In this way, the robot's skills can be automatically and iteratively improved in a cyclic process that combines simulation with reality.
In addition, the robot can track the opponent's behavior and playing style to adapt to different opponents, such as which opponent tends to hit the ball back to the table.
This allows you to try different techniques, monitor your success rate, and adjust your strategy in real time.
In experiments with humans, the team also discovered that the robot had a weakness: it was not good at handling downspin balls.
According to the estimation of ball rotation, the robot's play-up rate was plotted, and the results showed that when faced with more downspin balls, its play-up rate decreased significantly.
According to the researchers, the robot has a hard time handling balls with low arcs close to the table to avoid hitting the table, and there are limitations in determining the type of rotation of the ball in real time.
Studying bots to play table tennis started a long time ago. There is also a basket of team-related research:
For example, in Google's previous i-Sim2Real study, the trained robot can play ball with humans up to 340 times in a row without landing, which is equivalent to playing for 4 minutes + in a row.
Other teams have also had table tennis robots, such as this aunt's, which can also serve:
As well as teams like the Japan national team and Taiwan, China, also use robots to accompany their own Olympic athletes.
So some friends are puzzled, what is the difference with the robot released by Google this time?
Some netizens gave an explanation:
Google is about AI Agent this time, which works through video input, not pre-programmed algorithms.
So, when will we see a fight against our national team? （Doge）
This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com