Eka's robot achieves fluid dexterity by training in simulation, moving beyond prior systems that struggled with real-world object manipulation.

Eka, a Cambridge, Massachusetts startup cofounded by MIT professor Pulkit Agrawal and ex-Google DeepMind roboticist Tuomas Haarnoja, has built a robot arm that can pick up, manipulate, and handle diverse objects—including screwing a light bulb into a socket—with natural, continuous motion that prior robot arms could not achieve.

Unlike earlier approaches such as OpenAI's Dactyl (which could only manipulate a single instrumented Rubik's Cube under precise conditions), Eka's system trains robotic hands in detailed virtual simulations for thousands of computer hours, learning to handle objects of varying sizes, shapes, and weights without relying on human-provided training videos.

Agrawal frames robot dexterity as foundational: "Trillions of dollars flow through the human hand," suggesting applications beyond factories and warehouses to shops, restaurants, and households. The cofounders believe they are halfway to solving dexterity and say the remaining work is "just a question of scaling up the approach."