AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 20, 2026

Eka Robotics, a startup building a robotic arm and gripper for object manipulation, has closed a $150 million Series A round at a $1.45 billion post-money valuation. The round was led by venture funds a16z and F-Prime. The company was founded by Pulkit Agrawal, an MIT professor, and Tuomas Haarnoja, a former Google DeepMind researcher. The two first met in graduate school at the University of California–Berkeley, and both are known for work on reinforcement learning for robot control.

The pitch rests on a manipulation architecture that combines a force-sensitive gripper with a new kind of deep-learning model. Eka says its robot can pick up about 1,500 objects weighing up to 1 kg in an hour, against roughly 600 for a human, and that it reliably grasps objects more than 95% of the time. The arm also moves fast and with enough strength that it could damage itself, which one observer read as a sign of confidence in the control algorithm.

The most striking claim is generality without retraining. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, said the Eka robot adapted to a large number of different objects without being retrained, and described a "humanness" in how it would retry a failed grasp again and again. Another roboticist, who asked not to be named because of work at a frontier lab, compared the moment to the arrival of reliable legged locomotion five to 10 years ago, and described handing the robot a raspberry and watching a large, heavy machine slow down just in time to touch their arm gently. a16z general partner Raghu Raghuram, who co-led the round, said the system already exceeds superhuman performance on certain tasks and predicted that will extend to many dexterous manipulation tasks over the next few months.

The technical bet is on force rather than vision alone. Where most current systems are vision-language-action models, trained on the relationship between language, visual inputs and actions, Eka's is a vision-force-action model that brings force sensing into the loop. Agrawal's argument is that vision tells a robot roughly where an object is and how it is oriented, but the final millimeter of contact and the search for a stable grip happen through forces. That framing has a well-known constituency: critics of the generative-AI direction in robotics, including Rodney Brooks, have argued that models ignoring force and touch will fall short. Eka's results are being read as support for that position.

The hardware had to change to match. Eka designed its own hand, more flexible than a conventional gripper and carrying more force sensors, so it reacts to contact — the reason it can hold a raspberry without crushing it. The founders are guarded about the specifics, but the method is said to rest on a physics simulator that models the many forces encountered during a grasp, feeding a reinforcement-learning environment that can run at scale. Eka says it can generate thousands of hours of robot experience in minutes. That matters because it sidesteps the usual data pipeline of human demonstrations, teleoperation by gig workers, or long operator sessions. Hauser's read is that reinforcement learning lets the robot surpass whatever demonstrations it was given rather than being capped by them.

This is a meaningfully different starting point from much of the field. The dominant approach pushes high-level intelligence downward — a general model plans, and lower layers execute. Haarnoja describes Eka as working bottom-up, starting from physical competence. In practice the two are complementary: a general model might know that plates go from the dishwasher into a cupboard, while a manipulation model handles not breaking the dish on the way. Eka frames its contribution as closing the gap between generality and performance that dogs modern robotics, where the most general models attempt many tasks but do them poorly, and single-task specialists are too narrow to justify deployment.

The commercial plan is deliberately narrow. Eka is running undisclosed pilots with industrial customers in pick-and-place and wire manipulation, with the stated intent of making those two verticals succeed before widening scope. Cost and durability are part of the story: Agrawal says the gripper hardware costs less than $1,000 and is rated for 5 million duty cycles, against a target of 20 million, the bar set by standard industrial arms. Going from a lab demo to an assembly line is a substantial step; Hauser notes that running for a full year without failure alongside integrators will be a major challenge, and that the demonstrated work is pick-and-place, while many value-added industrial tasks involve more than that.

A reasonable reading is that Eka's strongest asset is not the arm but the simulation-plus-reinforcement-learning loop, if it holds up outside controlled conditions. Sim-to-real transfer is where such systems commonly degrade, and reliability rates measured on a demo bench say less about a year of uptime. The force-sensing gripper also sits in a cost and wear regime that industrial buyers will scrutinize closely. Still, the combination of sub-$1,000 hardware, high grasp rates without retraining and a stated path to industrial duty cycles is what drew a nine-figure round at a unicorn valuation.

Why it matters: Developers working on embodied AI now have a concrete alternative to vision-language-action stacks, one that puts force sensing and large-scale simulated reinforcement learning at the center. If Eka's grasp rates and retraining-free generality survive integration on real lines, teams building warehouse, assembly or wire-handling automation may need to reconsider whether imitation learning from human demonstrations is the right default. That remains an inference: the pilots are undisclosed and the durability targets are not yet met, so the near-term effect is likely pressure on data-collection strategies rather than an immediate shift in production floors.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Eka Robotics may be the first robotic arm that is actually good at picking up a wide range of objects, and it is so smooth and human-like, it is wowing the industry.