Eka Robotics raises US$150M at a US$1.45B valuation for robots that grip by feel

Most robot foundation models today are vision-language-action models: they look, read an instruction and move. Eka's is a vision-force-action model. It combines camera input with force feedback from a custom gripper, so a grasp that starts to slip can be corrected and resistance during an insertion changes the robot's motion instead of throwing an error[3].

"Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces," Agrawal told TechCrunch[2].

The other bet is how it is trained. Rather than relying on hours of human demonstrations, Eka uses reinforcement learning in a physics simulator, which it says generates thousands of hours of robot experience in minutes, then transfers that to real hardware[1]. It compares the approach to AlphaZero, which reached superhuman play in chess and Go without human data.

Eka says its robots pick a wide variety of objects outside their training distribution at about 1,500 operations an hour, roughly two and a half times a human baseline of about 600, and that the same robot can screw in a light bulb and insert wires[1]. TechCrunch adds that the objects weigh up to 1 kg and that Eka reports grasp success above 95%[2]. a16z is careful to say those are laboratory induction and bin-picking results, and that "sustaining that performance through full production shifts is the next test"[3].

Outside roboticists who have seen it are impressed. University of Illinois professor Kris Hauser told TechCrunch the robot "was able to adapt to a large number of different objects without retraining"[2]. Obvious Ventures says the robots work without safety cages and are gentle enough to take a single raspberry from a child's hand[4].

Eka is running pilots with industrial customers in two areas, pick-and-place and wire handling, but hasn't named them[2]. Its gripper costs under US$1,000 and is rated for 5 million duty cycles; Agrawal says the goal is the 20 million cycles expected of standard industrial arms[2]. The company says the money will go into extending the model to more tasks and speeding up deployments[1].

Dexterous handling is the gap between today's demo robots and ones that earn their keep on a factory floor. Large language models can increasingly plan what a robot should do; Eka is betting the hard part is the last millimetre of actually doing it, and that simulation plus touch gets there faster than collecting ever more human video. The open questions are the ones Hauser raises: whether it runs for a year on a real line without breaking, and how far it generalises beyond pick-and-place[2].