Eka Robotics valued at $1.45B after $150M round for touch-driven robot manipulation

Eka Robotics has closed a $150 million Series A at a $1.45 billion post-money valuation, according to TechCrunch, with a16z and F-Prime leading the round. The startup, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, is attacking one of robotics' oldest unsolved problems: getting machines to pick things up reliably, despite years of investment and mountains of training video.

A rare amount of buzz for a grasping demo

TechCrunch reports that Eka's demonstrations have startled veteran robot builders. Kris Hauser of the University of Illinois Urbana-Champaign described a distinct wow factor, noting that the robot adapted to many different objects without retraining and kept retrying after failures in a way that felt strikingly human. A roboticist at a frontier lab, speaking anonymously, said watching the arm work was emotionally comparable to the moment robots first achieved reliable legged locomotion.

The reported numbers support the enthusiasm. The robot can pick up roughly 1,500 objects per hour weighing up to 1 kilogram, compared with about 600 for a human, and Eka claims it succeeds more than 95% of the time. The arm moves fast enough to damage itself, which the anonymous observer interpreted as evidence of deep confidence in the control algorithm.

Touch, not just sight

Most current physical AI efforts build on vision-language-action models, which learn relationships among language, visual input and actions. Eka instead trains a vision-force-action model, treating force as the core signal for the final approach. Agrawal told TechCrunch that vision roughly locates an object, but the actual searching and grabbing happens through forces. That echoes a long-standing argument from critics such as Rodney Brooks, who have said generative robotics models will fail without representing touch.

The hardware matters as much as the model. Eka designed a flexible gripper with more force sensors than conventional designs, one that reacts on contact — which is why it did not crush a raspberry when a tester handed one over. The founders are guarded on specifics for competitive reasons, but the approach reportedly centers on a physics simulator that models the forces encountered during a grasp, feeding a reinforcement learning environment that runs at scale. The company says it can generate thousands of hours of robot experience in minutes, freeing it from dependence on human demonstration data.

The bottom-up complement to general models

Eka frames itself as the physical half of a two-layer stack. Haarnoja told TechCrunch that most of the field works top-down, starting from high-level intelligence and pushing toward the physical, whereas Eka starts from the physical layer and builds up. In practice, a general model might know that plates belong in a cupboard after a dishwasher cycle, while Eka's model handles lifting a dish without breaking it. Raghu Raguram, the a16z partner who co-led the round, told TechCrunch he expects superhuman performance across many dexterous tasks within months.

Still unproven outside the lab

Eka is running undisclosed pilots with industrial customers in pick-and-place and wire handling, but the jump from lab to production line remains large. Hauser flagged two open questions: whether the system can run for a full year on a real assembly line without breaking, and whether the technique generalizes beyond pick-and-place to more value-added industrial tasks. On hardware, the gripper — which costs under $1,000 — is currently rated for 5 million duty cycles, short of the 20 million cycle bar set by standard industrial arms. Agrawal says the company is deliberately narrowing its focus to two verticals before expanding, describing the ambition as building the engineering for entire workflows from a pragmatic starting point.

Why it matters

Manipulation, not locomotion or planning, is now the bottleneck for useful general-purpose robots. Eka's bet — that force sensing plus reinforcement learning in simulation can crack reliable grasping — directly challenges the vision-first orthodoxy of the current physical AI wave. If it holds up on real production lines, it supplies the missing physical layer that slow, expensive general planners need, and validates critics who argued touch was the overlooked ingredient. If it doesn't survive a year of industrial duty cycles, a $1.45 billion valuation will look premature. Either way, two major funds have now put serious money behind the idea that the last millimeter of robotics is a force problem, not a vision problem.