Five Years to Teach a Robot Table Tennis | Mireille El Geche, Sony AI
Mireille El Geche of Sony AI on reinforcement learning, physical AI research in Zurich and what a table tennis robot teaches us about moving AI from simulation into the real world.
Key insights: executive summary
- A table tennis robot is a benchmark for physical AI: perception, prediction and actuation all have to work inside a few hundred milliseconds.
- Reinforcement learning results in simulation transfer badly until the hardware, latency and sensing pipeline are treated as part of the model.
- Long-horizon research projects need corporate labs with Zurich-based teams that can absorb five-year timelines.
- Skills learned on a sports robot generalise to industrial physical AI: contact-rich manipulation, fast control loops and safety around humans.
- Zurich's research density means a specialist in one narrow control problem can find collaborators within walking distance.
“The gap between a simulated policy and a real rally is where all the actual engineering lives.”
“Five years on one problem is not slowness. In physical AI it is what depth costs.”
Why is a table tennis robot such a hard AI problem?
The ball moves faster than a comfortable perception budget allows, spin is only partially observable, and the robot must commit to a trajectory before it has full information. That forces prediction, not reaction, which is the same constraint industrial robots face when working next to moving humans.
What does this teach us about physical AI commercialization?
Robustness, not peak performance, decides whether a system ships. A policy that wins one rally in a lab is a demo; a policy that behaves predictably across lighting, wear and hardware drift is a product.
Who is Mireille El Geche?
Mireille El Geche is a Robotics and AI researcher at Sony AI, part of the Zurich deep tech ecosystem this series documents. Topics covered in this episode: Physical AI, Reinforcement learning, Robotics research Zurich, Sim-to-real transfer.
Visit Sony AIEpisode show notes
You have probably seen the videos of a robot rallying against a real table tennis player. This is the five years behind it. Mireille El Gheche was recruited to Sony AI with two research papers and no job description. She had to work out what the project was. She guessed right, and went on to lead a team of engineers and researchers on a problem nobody had solved. We talk about why a robot is so much harder than an AI system, what technical leadership actually demands, where robotics is heading, and whether she would trust a fully autonomous surgical robot with her own body.
Follow the Swiss tech ecosystem, one drive at a time
New in-car interviews with founders, investors and researchers building Zurich's deep tech, AI and robotics scene.