Robots often struggle with a deceptively simple job: picking up objects without dropping or crushing them. Eka Robotics says its combination of a force-sensitive hand and a new learning model can make that task quicker and more reliable. The company has drawn attention from researchers and investors, though its performance on factory lines is still an open question.
A robot that reacts to touch
Eka was founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja. The startup has closed a $150 million Series A round, led by a16z and F-Prime, at a $1.45 billion post-money valuation.
In demonstrations described by TechCrunch, the robot adapts to many objects without retraining and keeps trying after a failed grasp. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, called the demo a “wow factor” and noted its “creepy humanness” when it tried again.
Eka says its robot can pick up about 1,500 objects weighing up to 1 kg in an hour. The article compares that with about 600 for a human, and reports Eka’s claim that the robot succeeds more than 95% of the time. Its quick, forceful movements also raise a practical issue: the arm could hurt itself if its control system makes a mistake.
Why force matters for a careful grasp
Many robotics models are built around vision, language and action. Eka’s approach instead emphasizes vision, force and action. Vision can help the robot locate an object and estimate its pose, while sensing contact helps guide the final part of the grasp—the moment when a hand must make contact without slipping or applying too much pressure.
That difference is central to handling something fragile. Eka designed a more flexible robotic hand with additional force sensors, so it can respond to contact rather than simply close around an object. In the source article’s example, that sensitivity helps it avoid crushing a raspberry.
The founders describe using a physics simulator to model forces a robot may encounter while grasping. Those simulated interactions feed a reinforcement-learning environment that can run at scale. The company says it can generate thousands of hours of robot experience in minutes, offering a way to train beyond the demonstrations collected from people performing tasks.
That training strategy could address a familiar constraint: a robot may only learn what appears in its examples. Hauser said Eka’s use of reinforcement learning could let the robot go beyond the demonstrations provided to it. The broader bet is that simulation can make dexterous skills practical to train at scale, even though interacting with objects has historically been harder than avoiding them.
Physical skill could complement general AI
Robotics faces a balance between versatility and performance. Models that can attempt many tasks may not perform them especially well, while a system optimized for one task may not be useful enough to justify its cost. Eka’s proposed division of labor is to let a more general model decide what needs doing, while its model handles the physical act of doing it.
For example, a general model could determine that plates belong in a cupboard after a dishwasher cycle. Eka’s model would be responsible for picking up a plate without breaking it. Haarnoja describes this as starting “from the bottom up,” focusing on physical control rather than starting with high-level planning and pushing it down toward the robot.
The idea is distinct from a general model operating a robot in an unfamiliar setting, which the source says can be slow and costly. Eka is focused on the part that happens at the point of contact: grasping and manipulating objects reliably enough for a workflow to succeed.
Factory performance is the next test
Eka is working in secretive pilots with industrial customers on pick-and-place and wire manipulation. Those pilots are an early step, not proof that the system can work reliably on a production line. Moving from a simulation or lab demonstration into a full assembly workflow brings new demands, including operating for long periods without failure.
Hauser identified year-long operation without breaking as a key challenge. He also questioned how well the techniques will generalize beyond pick-and-place, since many industrial tasks involve more than moving an item from one location to another.
Hardware durability is another measure of readiness. Agrawal says Eka’s gripper costs less than $1,000 and is rated for 5 million duty cycles. The company’s stated goal is to reach the 20 million cycle bar set by standard industrial arms.
For now, Eka is concentrating on two areas: pick, arrange and place, and wire handling. If the system proves reliable in those settings, the approach could have room to expand. The funding and early demonstrations show strong interest; whether that promise translates into dependable factory work will depend on what the pilots can establish.