Five Years to Teach a Robot Table Tennis | Mireille El Gheche, Sony AI
Mireille El Gheche of Sony AI on reinforcement learning, the sim-to-real gap and the five-year effort to teach a robot to play table tennis.
Discuss this episode and meet other Swiss tech builders in the TechDrive WhatsApp group.
Key insights: executive summary
- The sim-to-real gap is robotics' hardest challenge. A model can reset endlessly in simulation, but real-world physics is unforgiving of even millisecond delays.
- Breakthrough robotics requires research scientists and hardware engineers to work in lockstep because algorithms built in isolation fail against physical constraints.
- As teams scale, technical leaders must deliberately move from individual contribution towards trust, delegation and management.
- Simulated data accelerates training and reduces the cost of collecting rare edge cases in the physical world.
- Representation in robotics gives the next generation of engineers stronger networks and the confidence to enter and lead in deep tech.
“If you build an AI system, it lives in the digital world... but if you move to the real world, the physics doesn't forgive you.”
“Building an AI system has to think. Building a robot has to think, survive the physics, and interact with the real world.”
What is the sim-to-real gap and why does it matter now?
The sim-to-real gap is the difference between how a robot performs in a digital physics engine and in the physical world. As companies race to deploy physical AI, closing it is essential to zero-shot transfer, where a robot trained in simulation can operate reliably in reality.
How can AI teams handle the complexity of real-world physics?
Use simulation for the heavy lifting of training, then use physical deployment for validation and fine-tuning. Researchers and engineers must work closely so theoretical models account for hardware latency and other constraints from day one.
Merge research and engineering
Researchers need direct knowledge of physical hardware constraints rather than working in a separate silo.
Find courage in community
Communities such as Women in Robotics provide networks, representation and confidence for people pushing boundaries in a male-dominated field.
Who is Mireille El Gheche?
Mireille El Gheche is a Senior Research Scientist at Sony AI and an advocate in Women in Robotics. Her work spans reinforcement learning, physical AI, sim-to-real transfer and interdisciplinary technical leadership.
Episode show notes
Mireille El Gheche describes Sony AI's demanding five-year journey to teach a robot to play table tennis with human players. She traces the project from small prototypes to a system capable of responding to the speed and spin of a real game.
What stood out was her perspective on leadership: the subtle and often difficult move from hands-on research to managing a team of ten, where building trust and clarity across disciplines becomes the central challenge.
More conversations from the Swiss tech ecosystem
- Nobody Has Left a Robot Working Overnight | Yulia Sandamirskaya, ZHAWYulia Sandamirskaya, ZHAW Zurich University of Applied Sciences
- Physical AI Is Ready, Buildings Are Not | Kateryna Portmann, ANYboticsKateryna Portmann, ANYbotics
- Switzerland's AI Movement Is About to Start | Swiss {ai} WeeksKatka Letzing and Anna Hakobyan, Swiss {ai} Weeks
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. Want to discuss this episode, ask follow-up questions and meet other builders?