Eika valued at $1.45 billion: robot picks 1,500 pieces per hour using force sensing

Grasping objects remains the biggest weakness in robots, despite billions of dollars spent and endless videos that have fed deep-learning models. Eika Robotics bets that the cause is not a lack of data but a misreading of what happens in the last millimeter before the piece is grasped.

The company closed a Series A round of $150 million, valuing it at $1.45 billion after the round, led by the a16z and F-Prime funds. It was founded by Polciet Agrawal, a professor at MIT, and Thomas Haraina, a former researcher at Google DeepMind.

The robot picks up about 1,500 pieces per hour, each weighing up to a kilogram, compared with roughly 600 for a human.The company says it succeeds in picking up more than 95 percent of the time.

The secret lies in the model architecture. Instead of the usual vision-language-action model, Eika built a vision-force-action model, linking language, force perception and motion. Agrawal says vision tells you roughly where the object is and its pose, while the actual grasp is decided by force. Therefore the robot hand was designed to be more flexible than a conventional gripper, equipped with more force sensors that respond to contact, so it does not crush a berry.

The training itself defies the norm. The company uses a physics simulator to model the forces the robot will encounter, builds a large-scale reinforcement-learning environment on top of it, and says it generates thousands of hours of robot experience in minutes. Chris Hauser, a professor at the University of Illinois Urbana-Champaign, sees this as freeing the robot from the limits of recorded human demonstrations. He said the robot adapts to a large variety of different objects without retraining, and it retries if it fails the first time.

The picture is not entirely rosy. The company is conducting secret trials with industrial customers on part transport, sorting and wire handling, but Hauser warns that the real challenge will come when it works with integrators on actual assembly lines that must run continuously for a full year without failure. He also questions whether these technologies can move beyond picking and placing to other industrial tasks. The company says it is starting with only two domains, and that its gripper, costing less than $1,000, is rated for five million cycles, with a goal of reaching 20 million cycles, the benchmark for industrial arms.

Here the impact begins for those working in the Gulf, Egypt or the Levant, a deduction from the source’s facts rather than a regional report. If it is true that a gripper costing under $1,000 can perform this task, the feasibility question in warehouses and light packaging-assembly plants shifts from “can the robot even do it?” to “how many cycles can it endure and what is its maintenance cost.” A procurement manager evaluating automation will need to ask for long-term reliability figures, not short, flashy proposals.

The required skill set is also shifting. The source describes a division of labor between a general model that decides what to do and a model that handles the physical side. This means that the integration engineer who connects the two on a real production line becomes a critical link, a skill that is built locally before being imported. Regional robotics startups now face a new competitor with deep pockets, yet they see a clear example that simulation can replace costly human data collection.

Caution remains warranted. The source says Eika still needs to prove itself in the field, and the gap between the lab and the production line almost equals the gap between simulation and reality. The prudent decision now is to continue with limited trials, not to rebuild an entire automation plan around a single offering, regardless of what experts may say about it.