Eka's robot picks 1,500 objects an hour in demos, vs ~600 for a human

A roboticist at a frontier AI lab, who asked not to be named, described the moment a big, heavy robot rushed toward their arm, slowed down just in time and touched them very gently. The robot had been handed a raspberry, the lab's apparent favourite test item, and it did not crush it. Robots have been learning from filmed demonstrations for years.

The startup behind it, Eka Robotics, was founded by MIT professor Pulkit Agrawal and Tuomas Haarnoja, a former Google DeepMind researcher. Its latest round, a $150 million Series A, values the company at $1.45 billion post-money, meaning after the new money is added. The round was led by the venture funds a16z and F-Prime. Raghu Raguram, the a16z general partner who co-led it, told TechCrunch he expects “superhuman performance in many dexterous manipulation tasks” within “the next few months”. That is an investor’s forecast, not a measured result.

The technical choice sits at the centre of the story. Most robot models pair vision with language, a design known as vision-language-action. Eka’s foundation model pairs vision with force, the push and pull a hand feels when it touches something. Agrawal put it plainly: “Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces.” The hand itself was redesigned with more force sensors so it can react the instant it meets resistance. The company trains that behaviour in a physics simulator, which models the forces a robot meets when it tries to grasp something, and runs the resulting reinforcement learning, a method where a program learns by trial and reward, at scale. Eka says this produces thousands of hours of experience in minutes, instead of the slow collection of footage by gig workers or operators. The founders have not disclosed how they tuned the simulation, citing competition.

What this changes, if the figures hold, is the division of labour between machines. A general model can plan a task, while a model like Eka’s handles the physical contact. Its speed has a cost: the robot arm is strong enough to damage itself as it moves fast, and Eka’s reliability figure of more than 95% is the company’s own. The reporting gives no date for work outside demonstrations, no customer and no price. For robot builders, the next test is whether the grip still holds in a factory or warehouse, on objects the system has never seen.