E-Paper
Why is Google making two robots play endless table tennis? The reason reveals the future of AI
At Google DeepMind, two robotic arms play endless table tennis matches as part of a project to help robots learn real-world skills.
At a lab south of London, two robotic arms been playing table tennis non-stop, pushing each other to new limits and quietly hinting at the future of artificial intelligence in the real world. Unlike the legendary Wimbledon marathon where humans finally called it quits, these robots seem content to keep going, always learning, never truly finished.
Training robots, one rally at a time
Google DeepMind’s project started as a hunt for better ways to train robots to handle real-world complexity. After all, it isn’t enough for a robot to just lift a box if it cannot adjust to unexpected changes or interact with people around it. The team decided that table tennis, a game that mixes fast reaction times, precision control, and strategic play, was a natural choice for testing. Every point, with its wild spins and shifting speeds, is a lesson in adapting to a moving target.
The first step was simple rallies. The robots played cooperatively, just keeping the ball in play. Gradually, engineers turned up the challenge, tweaking the rules so that each arm began to compete for points. Quick improvement wasn’t immediate; the robot arms forgot some tactics as fast as they learned new ones, and early rallies were often short and awkward. Progress ramped up, though, when real humans jumped in. Facing off against people with different styles, the robots began seeing a broader set of shots, forcing them to adjust and respond on the fly. After dozens of matches, these arms could routinely outplay beginners and even break even with some intermediate players.
What really sets this project apart is how the robots are now getting feedback. Google’s Gemini vision-language model watches clips of table tennis games, then gives clear, actionable advice: hit farther right, go for a short ball, defend closer to the table. Unlike old-school programming, this feedback comes in natural language, almost like a coach at the sidelines. The robots adjust their strategies and keep growing, rally by rally.
Why it matters beyond the table
There’s a bigger dream behind this marathon. DeepMind hopes that robots learning from endless competition and human coaching will one day lead to machines ready for real jobs. It’s a step toward robots as office helpers, lab partners, or just reliable hands in unpredictable home environments. In the world of robotics, mastering “simple” actions, like tying a shoelace or avoiding trip-ups, remains the real challenge, not chess or code-breaking. Long rallies at the table may help smooth that learning curve and chip away at obstacles that have slowed progress for years.
Researchers say these games are just the beginning. As AI models become more general and feedback loops tighter, the journey from lab-bound robot to everyday helper could speed up. Until then, the arms keep at it, never tiring, always volleying, and inching closer to a day when robots truly join us in the rhythm of daily life.
