Man VS Machine: Who Plays Table Tennis Better? 🤖

TL;DR

The robot learned to play table tennis quite well, returning 98% of balls after only 1.5 hours of training. It first practiced in simulation, then underwent quick retraining with a ball-throwing machine before facing a real player. Most returns landed within 25 centimeters, or about 10 inches, of the desired spot, though backspin remained a challenge. Read on to see how simulation enabled this rapid real-world performance.

Transcript

Dear Fellow Scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér. Today we are going to see if a robot can learn to play table tennis. Spoiler alert, the answer is yes, quite well in fact. That is surprising, but what is even more surprising is how quickly it learned to do that. Recently, we have seen a growing number of techniques whe... Read More

Key Insights

• 🤖 Robots can learn complex tasks quickly by training in simulated environments.
• 👻 Simulation training allows for safer and more efficient learning for tasks like driving or playing sports.
• 🤖 Automatic domain randomization enables robots to learn general knowledge from different variations of a problem.
• 🤖 The success of the table tennis-playing robot shows that impressive results can be achieved without extensive resources.
• 🤖 The robot's inability to handle backspin balls highlights the challenges that still exist in robot learning.
• 🎮 The limited viewership of this video emphasizes the importance of sharing and discussing these technological advancements.
• 💦 This work demonstrates the potential for future advancements in robotics and AI.

Questions & Answers

Q: Can a robot learn to play table tennis?

Yes. After practicing in simulation and completing a quick retraining step against a ball-throwing machine, the robot successfully played table tennis against a real player.

Q: How long did the robot take to learn table tennis?

The robot took only 1.5 hours of training. After this training, it was able to transfer what it had learned into the real world.

Q: How accurately could the table tennis robot return the ball?

The robot returned 98% of the balls. Most of its returns landed within 25 centimeters, or about 10 inches, of the desired spot.

Q: How was the table tennis robot trained?

The robot first practiced in a simulation where the ball was sent to its side and a desired return position was specified. It then completed a quick retraining step against a ball-throwing machine before playing in the real world.

Q: Why is table tennis difficult for a robot to learn?

Playing table tennis requires finesse, rapid movement, and prediction of what will happen in the near future. Spin also makes the ball’s trajectory harder to predict.

Q: Which types of spin could the robot handle?

The robot practiced against side-spin and top-spin balls. It could not yet handle backspin because it caused too much acceleration in a robot joint.

Q: What is automatic domain randomization?

Automatic domain randomization creates many random environments that differ slightly from one another. The AI learns to solve multiple variants of the same problem so it can develop more general knowledge.

Q: What other systems learned in simulation before real-world deployment?

OpenAI’s robot hand learned in simulation to rotate a Rubik’s cube to a target state using automatic domain randomization. Tesla also uses a simulated game world to train self-driving cars safely across varied “what if” scenarios.

Summary & Key Takeaways

• Robots can quickly learn complex tasks like playing table tennis by training in simulated environments.

• OpenAI's robot hand and Tesla's self-driving cars have successfully learned in simulation before being deployed in the real world.

• The table tennis-playing robot achieved a 98% return rate and accurate ball placement within 25 centimeters of the desired spot after only 1.5 hours of training.