6 points | by aanet 6 hours ago ago
2 comments
Sony AI's "Ace" robot outperforms humans in table-tennis (ping-pong)
Archive: https://archive.is/20260422165419/https://www.ft.com/content...
Paper at Nature: https://www.nature.com/articles/s41586-026-10338-5
Utterly fascinating to see how the Ace robot from Sony AI was built to play with and beat elite level ping-pong players.
Video here: https://youtu.be/EH8kZDc7OLk?si=YAnQi-lWE28QIkVA
A few things caught my eye:
- Perception system tracks ball's spin at over 150 revolutions/second
- Top speed of the ball can exceed 20-25 meters/second; (the table is only 2.7m long)
- Reaction time of the robot can be as short as 20 milliseconds, roughly 10x faster than elite athletes
- Event cameras used to detect the spin of the ball; regular cameras at 200Hz to detect the ball's trajectory in realtime (!)
- Robot trained in simulation-only first using RL techniques, then transfers learning to real world
Fascinating research!
Sony AI's "Ace" robot outperforms humans in table-tennis (ping-pong)
Archive: https://archive.is/20260422165419/https://www.ft.com/content...
Paper at Nature: https://www.nature.com/articles/s41586-026-10338-5
Utterly fascinating to see how the Ace robot from Sony AI was built to play with and beat elite level ping-pong players.
Video here: https://youtu.be/EH8kZDc7OLk?si=YAnQi-lWE28QIkVA
A few things caught my eye:
- Perception system tracks ball's spin at over 150 revolutions/second
- Top speed of the ball can exceed 20-25 meters/second; (the table is only 2.7m long)
- Reaction time of the robot can be as short as 20 milliseconds, roughly 10x faster than elite athletes
- Event cameras used to detect the spin of the ball; regular cameras at 200Hz to detect the ball's trajectory in realtime (!)
- Robot trained in simulation-only first using RL techniques, then transfers learning to real world
Fascinating research!