Eka is worth $1.45B thanks to a physical AI model that has experts raving

Eka Robotics, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, has closed a $150 million Series A at a $1.45 billion post-money valuation, co-led by a16z and F-Prime. That is a big number for a company still running secretive industrial pilots, and it rests on one claim: that Eka has cracked reliable, fast robotic grasping.

The reported figures are concrete. Eka says its system picks up about 1,500 objects of up to 1 kg per hour, against roughly 600 for a person, and succeeds more than 95% of the time. Outside researchers who saw it in person, including UIUC professor Kris Hauser, described a rare 'wow factor' and adaptation to many objects without retraining. Picking things up is the unglamorous bottleneck that has kept a lot of warehouse and factory automation stuck, so a real step here would matter far beyond one startup.

We find this one exciting, and for a specific reason: Eka is not just scaling the current playbook, it is arguing with it. Most of the field is building vision-language-action models trained on mountains of video and human teleoperation. Eka built what it calls a vision-force-action model and a gripper packed with force sensors, on the premise that the last millimeter of a grasp is a matter of touch, not sight. Critics like Rodney Brooks have long said generative robotics would stall without representing force and touch. If Eka's results hold, those critics were at least partly right, and that is healthy for a field prone to herd behavior.

The training approach matters as much as the hardware. Instead of leaning on video collected by gig workers or operators, Eka models contact forces in a physics simulator and runs reinforcement learning at scale, claiming thousands of hours of robot experience in minutes. Hauser's point is the key one: a robot that learns by trial in simulation can exceed the demonstrations it was given, rather than imitating them. That is the path that cracked legged locomotion, and one roboticist compared seeing Eka's arm to that moment.

Now the skepticism we owe readers. The founders are deliberately vague about how they built it, everything we know comes from demos and the company's own numbers, and the pilots are not public. Demos are not production. Hauser named the real test: running on an integrator's assembly line for a full year without breaking, and doing more than pick-and-place. The gripper is rated for 5 million duty cycles today, against the 20 million that standard industrial arms meet, so durability is an open engineering problem, not a solved one.

There is also a physical-safety thread worth naming. The arm is strong and fast enough to damage itself, and one observer described it rushing at their arm before slowing to a gentle touch. That is impressive control, and it is also exactly the kind of behavior that has to be proven fault-tolerant before these systems work near people outside a lab. Trusting the algorithm not to break the robot is one thing; trusting it not to hurt a worker after a sensor fault is another.

And the labor question is unavoidable. A machine that out-picks a human by more than two to one, with hardware under $1,000, is a direct line to fewer warehouse and line jobs if it scales. We think better manipulation is net useful — it can take over dull, injury-prone work — but the gains will not land evenly, and the people doing pick-and-place today are the ones carrying the risk.

Agrawal and Haarnoja met as graduate students at UC Berkeley and are known for reinforcement learning research in robot control. Eka is starting with two verticals — pick, arrange and place, plus wire handling — and frames its model as the physical layer that a general planning model could sit on top of, handling the grip while a larger model decides what to do.