In a funding round led by Andreessen Horowitz and F-Prime, the Cambridge-based startup Eka Robotics announced a $150 million Series A investment that lifts its post-money valuation to $1.45 billion. The capital injection follows an $11 million seed round raised in May 2025 and comes less than a year after the company’s founding by MIT professor Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja. With roughly 25 employees on the payroll, the valuation translates to about $40 million per staff member, a metric that underscores the market’s confidence in the firm’s breakthrough approach to robotic manipulation.
How the Vision-Force-Action model reshapes robot hands
Eka’s flagship robot is built around a proprietary Vision-Force-Action (VFA) foundation model. Unlike typical vision-language-action pipelines that rely primarily on visual cues, the VFA model fuses camera data with real-time force feedback from a custom-designed, highly flexible gripper. This force-sensitive gripper houses an array of tactile sensors that allow the robot to feel the moment it contacts an object, adjusting its grip before damage occurs. The result is a system that can manipulate items ranging from delicate berries to 1-kilogram metal blocks with a success rate exceeding 95 %.
Performance metrics released by the company illustrate the speed advantage over human workers. The robot can handle roughly 1,500 objects per hour, compared with about 600 items per hour for a typical human picker. The arm moves so quickly that its own structure must endure significant inertial forces, a testament to the confidence the engineers have in their algorithmic safety margins. Observers at leading robotic labs have described the experience of the robot gently slowing as it approaches a hand-held raspberry as “remarkably human-like” and emotionally resonant.
From simulation to reality: reinforcement learning at scale
Rather than relying on extensive human demonstration data, Eka trains its manipulation policies through simulation-based reinforcement learning. By constructing a high-fidelity physics simulator that models contact dynamics, friction, and compliance, the virtual robot can practice millions of variations of object placement, surface texture, and slip conditions in a matter of minutes. This approach sidesteps the bottleneck of human-generated datasets, allowing the system to encounter failure modes that are rarely shown to a learning robot in real-world trials.
Both founders have deep roots in the field: Agrawal pioneered computational sensorimotor learning at MIT’s CSAIL, while Haarnoja co-created the Soft Actor-Critic algorithm at Berkeley and later refined it at DeepMind. Their combined expertise enables Eka to shrink the notorious sim-to-real gap, delivering policies that transfer reliably from the virtual sandbox to physical hardware. As a result, the robot can recover from variations—such as a shifting connector or a slipping grip—without requiring manual retuning by engineers.
Economic stakes and future deployment scenarios
The human hand remains the most versatile tool in the global economy, underpinning sectors from electronics assembly to food processing. While industrial robots dominate highly structured lines, the majority of manual tasks involve unstructured variability that still resists automation. By achieving reliable dexterous manipulation Eka promises to extend robot uptime into those high-mix, low-volume operations where labor shortages are already acute.
Investors view the technology as a gateway to a multibillion-dollar market. If the VFA model can be packaged for logistics warehouses, confectionery factories, or medical device assembly lines, the upside could be exponential. The next milestone for the company will be large-scale pilot programs that test the robot across thousands of real-world picks, providing the data needed to prove that the simulation-first learning pipeline can sustain commercial workloads.
