Andreessen Horowitz is leading the Series A in Eka Robotics, a physical AI startup developing a Vision-Force-Action model for robotic manipulation, the venture firm said in an announcement on October 20, 2026. The post, authored by partner Yoko Li and managing partner Raghu Raghuram, lays out Eka’s approach to dexterous manipulation and the firm’s rationale for backing the company.
What Eka Robotics Is Building
Li and Raghuram wrote that much of the world’s physical economy, from manufacturing and electronics assembly to logistics and food production, still depends on people using their hands, even as industrial robots have transformed highly structured production where fixtures hold every part in place. Large categories of physical work remain hard to automate because the real world is variable: parts move, materials deform, objects slip, and connectors do not align perfectly. The investors said they believe solving dexterous manipulation is one of the largest opportunities in robotics, and noted that labor is already a constraint across many of these industries.
On its website, Eka says it is building “intelligence for the physical world in its native language: force.” The company says robotics has until now required choosing between generality and performance, and describes its Vision-Force-Action model as a foundation that unites generality, performance, and safety. Eka presents three pillars for the technology: a single model that generalizes across objects, tasks, and environments; fast, reliable robots that adapt to the unexpected; and safe, collaborative, environment-agnostic operation alongside humans.
How the Vision-Force-Action Model Works
According to the a16z post, Eka has built a Vision-Force-Action model that combines visual perception with force feedback from its robotic hands, and it designs the model, sensing, controller, and hardware together so the robot can use physical contact itself as information. The post describes the resulting behaviors: a grasp that starts to slip can be corrected, contact with a surface can help locate an object, and resistance during insertion can change the robot’s motion rather than simply trigger an error. The stated goal is to preserve performance as conditions change without requiring an engineer to retune the robot for every material, object, or fit.
The announcement grounds that approach in the mechanics of the human hand, which it describes as having roughly two dozen degrees of freedom and thousands of specialized sensory receptors that detect sustained pressure and shape, light touch, the first signs of slip, high-frequency vibration, and the stretching of skin. Because rebuilding such a biological sensing system receptor by receptor would add enormous complexity, cost, wiring, compute, and potential points of failure, the post frames the engineering challenge as giving machines enough physical information, and intelligence, to achieve the same outcome with a much simpler sensing and actuation system.
Learning in Physics Simulation
Eka trains its manipulation policies through reinforcement learning in physics simulation, where virtual robots can practice millions of variations in object position, friction, and contact before a policy ever reaches physical hardware, according to the announcement. Li and Raghuram contrasted that approach with learning from human demonstrations, which they said consumes human time and struggles to convey useful information about all the ways a task can fail. They added that, if the approach works at scale, robot experience can increasingly scale with compute rather than human labor.
The post acknowledges the historical obstacle to that approach: contact dynamics are notoriously difficult to model, and small errors in friction, compliance, or actuator behavior can turn a successful simulated policy into a useless real-world one. It describes closing the sim-to-real gap as one of robotics’ hardest problems, and says Eka’s founders have spent much of the last decade attacking that problem from complementary directions.
Reported Performance and the Founding Team
Eka reports up to 1,500 picks per hour in laboratory induction and bin-picking tasks, two and a half times faster than humans, according to the announcement. The post states that the underlying system has demonstrated manipulation across fragile objects, slippery industrial parts, and contact-rich insertion tasks, and it identifies sustaining that performance through full production shifts as the next test. If those capabilities transfer reliably across production tasks, the investors wrote, manufacturers could introduce new products without starting the learning process from scratch, and tasks that were never worth a dedicated automation project would become addressable.
Co-founder Pulkit Agrawal of the Improbable AI Lab at MIT CSAIL is described in the post as one of the leading researchers in robot learning and a pioneer in computational sensorimotor learning, with work spanning reinforcement learning, force-aware control, dexterous manipulation, and moving policies from simulation into the real world. Co-founder Tuomas Haarnoja co-created Soft Actor-Critic at Berkeley, which the post calls one of the foundational algorithms in modern reinforcement learning, and later worked at Google DeepMind training complex humanoid behaviors in simulation and deploying them on real machines. The investors wrote that the founders are unusually well suited to closing the sim-to-real gap and have been demonstrating that it is shrinking.
Eka says its broader team comes from MIT, Berkeley, Harvard, CMU, BU, UPenn, DeepMind, Microsoft, and Boston Dynamics, and describes its researchers as pioneers of self-supervised learning, large-scale robot learning, world models, reinforcement learning, and sim-to-real robotics.
Credit: Source link
