Despite years of heavy investment and enormous datasets fed into deep learning systems, robotic manipulation remains one of the hardest unsolved problems in the field. Robots, for all their advances, still struggle with the basic act of picking things up reliably. A startup called Eka Robotics believes it has found a way to change that, combining a force-sensitive gripper with a novel deep-learning architecture that is already drawing serious attention from experts across the industry.

The company was co-founded by Pulkit Agrawal, a professor at the Massachusetts Institute of Technology, and Tuomas Haarnoja, a researcher previously at Google DeepMind. Together, they have built what they describe as a bottom-up approach to physical AI - one that starts with the mechanics of touch and force rather than high-level reasoning. The company has just closed a $150 million Series A funding round, placing its post-money valuation at $1.45 billion. The round was led by venture capital firms Andreessen Horowitz (a16z) and F-Prime Capital.

Raghu Raguram, the a16z general partner who co-led the investment, said the startup is already outperforming humans on certain tasks. "They are already exceeding superhuman performance for certain tasks, and so we think over the next few months there will be superhuman performance in many dexterous manipulation tasks," he said.

A Reaction That Surprised Veteran Roboticists

In a field where polished demonstration videos are commonplace and often misleading, Eka's system has managed to impress people who are difficult to impress. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, described the system as having a clear "wow factor" compared to other robot demonstrations he had recently observed. What stood out to him was the robot's ability to handle a wide variety of objects without needing to be retrained for each one. He also noted an almost unsettling quality to its persistence - a kind of "humanness, creepy humanness," in that the robot would continue attempting a task if it failed on the first try.

A second roboticist, who works at a leading AI lab and requested anonymity due to the sensitivity of their position, described the experience of watching the Eka robot in action as genuinely emotional - comparable, they said, to the moment years ago when robots first achieved consistent and reliable legged locomotion. That milestone, widely regarded as a turning point in the field, took place roughly five to ten years ago.

What makes the robot's performance notable is not just its success rate but its speed and fluidity. The system can handle approximately 1,500 objects weighing up to one kilogram within a single hour - more than double the roughly 600 objects a human worker could process in the same time. Eka reports a successful pick rate above 95 percent. The robot arm operates at speeds fast enough that it could theoretically damage itself through its own movements, a detail that the frontier lab observer found telling. "They have a lot of trust in their algorithm not breaking the robot," the observer noted. "I wonder how many robots they broke along the way?"

During a meeting with the Eka team, the same roboticist handed the robot a raspberry - apparently a standard test item at their lab. The encounter left a lasting impression. "The experience of a big, heavy robot rushing towards my arm and then slowing down just in time and very gently touching me is a feeling that I am, as we speak, carrying this feeling in my body," they said.

Why Force Sensing Changes Everything

The performance Eka has achieved traces back to a fundamental architectural decision made during the design of its foundational model. Most robotics AI systems today are built around what is called a vision-language-action model - a framework that trains the system to understand the relationships between language commands, visual inputs, and physical actions. Eka took a different path, building what it calls a vision-force-action model, which centers the relationship between visual data, force inputs, and physical actions.

This distinction matters enormously in practice. Critics of mainstream generative AI robotics approaches - including prominent voices like Rodney Brooks, a pioneer in the field - have long argued that systems relying purely on vision will fall short when it comes to real-world manipulation, where the final millimeters of contact require something more than a camera can provide. Eka's results appear to support that critique.

"Vision is not the primary way in which you close this last millimeter problem. Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces."

To make this approach work, Eka had to design a custom robotic hand that is more flexible than conventional grippers and incorporates a greater number of force sensors. The hand responds dynamically to physical contact - which is precisely what prevents it from crushing a delicate object like a raspberry. While the founders are understandably guarded about the specific engineering details, the underlying method involves using a physics simulator to model the full range of forces the robot will encounter during a grasping attempt. That simulation environment is then used to build a reinforcement learning framework that can be scaled up dramatically. According to the team, the system can generate thousands of hours of robot experience within just a few minutes of compute time.

Hauser highlighted why this approach is significant from a research perspective. "Rather than being limited to the data that you're collecting from, say, human demonstrations, Eka is pushing through reinforcement learning this approach where the robot can surpass any kind of demonstrations that are given to it," he said. "What's really exciting is that they're able to not be limited to the data that is presented to them for training."

Simulation as the Path to Scale

Agrawal and Haarnoja first crossed paths during their graduate studies at the University of California, Berkeley, and both have built careers around reinforcement learning for robotics control. The central thesis behind Eka is that this technique can be executed at scale through simulation, removing the need to rely on video data collected by human workers or robot operators manually performing repetitive tasks over extended periods.

Simulation-based training has historically been more tractable for tasks that involve navigating or avoiding objects than for tasks that require physically interacting with them. Manipulation - grasping, placing, handling - has proven particularly resistant to the sim-to-real transfer that makes simulation useful. Eka's claim is that it has found a way to make that transfer work reliably for manipulation tasks.

"What Eka is bridging is the generality-performance gap, which plagues a lot of modern robotics," Agrawal said.

Filling the Gap Between General and Capable

The current state of physical AI presents a difficult trade-off. The most broadly capable models - those designed to attempt a wide variety of tasks - tend to perform each individual task at a relatively low level of quality. Highly specialized models, on the other hand, may be excellent at a single task but fail to justify the cost of deployment across an industrial operation that requires flexibility. The commercial promise of generative AI in robotics has always been that robots could learn new tasks quickly, including ones they had not explicitly been trained for.

Existing general-purpose approaches have made progress on this front. Systems like Google's Astra, built on top of capable large language models, can direct a robot through unfamiliar environments and perform a range of tasks - but they do so slowly and at significant computational expense. The model Eka is developing is intended to complement this kind of high-level reasoning by handling the physical layer with precision and reliability. The division of labor, as the founders describe it, is that a general model like Astra might understand that clean dishes belong stacked in a cupboard, while Eka's model provides the physical dexterity to pick up a plate without breaking it.

"Most of the field has been focusing on the planning," said Haarnoja. "So they approach from the top: high-level intelligence, pushing that down towards the physical. Whereas we think about it from the bottom up."

From Lab to Factory Floor

Despite the early results, Eka still has significant ground to cover before it can claim success in real-world industrial settings. The company is currently running confidential pilot programs with a number of industrial customers, focused on pick-and-place operations and wire manipulation tasks. These are meaningful first applications, but the path from a controlled lab environment to a functioning production line is rarely straightforward.

Hauser identified two key challenges ahead. The first is operational durability. "What's going to be a big challenge is when they work with integrators on real assembly lines, and it needs to run without breaking for an entire year," he said. The second is breadth of capability. "The other open question is how generalizable these techniques can be to different tasks. What I've seen is pick and place, but many industry tasks that are value-added involve more than just pick-and-place."

Agrawal is already thinking about the hardware side of long-term deployability. The current version of Eka's gripper costs under $1,000 and is rated for five million duty cycles. The goal, he said, is to eventually reach the twenty million cycle threshold that is standard for conventional industrial robotic arms - a bar that reflects the kind of continuous, uninterrupted operation that manufacturing environments demand.

The company's near-term strategy is deliberately focused. Rather than trying to prove the technology across every possible application at once, the team has chosen two specific verticals - pick, arrange, and place operations, and wire handling - with the intention of demonstrating clear, undeniable results in those areas before expanding further.

"What we are building is that engineering of how to design entire workflows, but we are starting off being very pragmatic, and saying here are two verticals in pick, arrange, place, and wire handling, let's make them be home runs. Once we have made home runs on them, the sky is the limit in how much we can do with the approach we are developing."

The $150 million Series A gives Eka the resources to pursue those goals with serious intent. Whether the jump from impressive lab demonstrations to reliable, scalable industrial deployment follows the trajectory its backers and founders are projecting remains to be seen - but the technical foundation, and the reaction it has drawn from people who know the field well, suggests this is a company worth watching closely.