MIT researcher shows how robots learn physical tasks through simulation and practice

MIT associate professor Pulkit Agrawal, cofounder of Eka Robotics, discussed how robots differ from humans in learning physical skills. While watching tennis does not teach you to play, robots similarly need hands-on practice to translate planning into reliable execution.

Agrawal demonstrated Eka's approach to robotic manipulation, including delicate handling of objects like raspberries, achieving a picking rate of 1,500 objects per hour. The key insight: robots require speed, precision and reliability to become practical products.

Eka's systems were trained entirely in simulation using reinforcement learning—trial and error—before their skills transferred to physical robots. This simulation-first approach enables robots to learn without needing massive amounts of real-world data collection.