Key Takeaways
- • Andreessen Horowitz (a16z) is leading the Series A round in Eka Robotics, a physical AI startup, according to an announcement dated October 20, 2026 by partner Yoko Li and managing partner Raghu Raghuram.
- • Eka is developing a Vision-Force-Action model for robotic manipulation that fuses visual perception with force feedback from robotic hands, targeting dexterous tasks in manufacturing, electronics assembly, logistics, and food production.
- • Eka reports up to 1,500 picks per hour in laboratory induction and bin-picking tasks, about two and a half times human speed, per the announcement; sustained performance through full production shifts is the stated next test.
- • Eka trains its policies with reinforcement learning in physics simulation, a sim-to-real approach a16z says could let robot capability scale with compute rather than human demonstration labor.
- • Co-founders Pulkit Agrawal (MIT CSAIL Improbable AI Lab) and Tuomas Haarnoja (co-creator of Soft Actor-Critic, former Google DeepMind) anchor a team drawn from MIT, Berkeley, Harvard, DeepMind, Microsoft, and Boston Dynamics.
Summary
Andreessen Horowitz is leading the Series A financing in Eka Robotics, a physical AI startup developing what it calls a Vision-Force-Action model for robotic manipulation. The firm disclosed the investment in an announcement on October 20, 2026, authored by partner Yoko Li and managing partner Raghu Raghuram. The investors argue that much of the physical economy, including manufacturing, electronics assembly, logistics, and food production, still relies on human hands, because industrial robots have excelled only in highly structured settings where fixtures hold every part in place. They view dexterous manipulation as one of the largest opportunities in robotics, citing labor constraints across these industries.
According to the a16z post, Eka's model combines visual perception with force feedback from its robotic hands, and the company designs the model, sensing, controller, and hardware together so that a robot can use physical contact as information. The post cites examples including correcting a slipping grasp, locating an object through surface contact, and adjusting motion during insertion when resistance is encountered. Eka's stated goal is to maintain performance as materials and conditions change without engineers retuning the robot for each application. The company frames its approach around three pillars: a single model that generalizes across objects, tasks, and environments; fast and reliable adaptation to unexpected conditions; and safe, collaborative operation alongside humans.
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 deployment on hardware. Li and Raghuram contrasted this with learning from human demonstrations, which they described as time-intensive and poorly suited to conveying the full range of failure modes. The post acknowledges that contact dynamics are difficult to model and that small errors can render a simulated policy ineffective in the real world, identifying the sim-to-real gap as one of robotics' hardest problems.
Eka reports up to 1,500 picks per hour in laboratory induction and bin-picking tasks, roughly two and a half times human throughput, according to the announcement. The company says the system has demonstrated manipulation of fragile objects, slippery industrial parts, and contact-rich insertion tasks. Sustaining these results across full production shifts is identified as the next milestone. Co-founder Pulkit Agrawal, of the Improbable AI Lab at MIT CSAIL, is described as a leading researcher in robot learning, and co-founder Tuomas Haarnoja co-created Soft Actor-Critic at Berkeley and later worked at Google DeepMind on simulation-trained humanoid behaviors. Eka says its broader team includes researchers from MIT, Berkeley, Harvard, CMU, BU, UPenn, DeepMind, Microsoft, and Boston Dynamics. The announcement did not disclose the Series A round size or valuation.
According to the a16z post, Eka's model combines visual perception with force feedback from its robotic hands, and the company designs the model, sensing, controller, and hardware together so that a robot can use physical contact as information. The post cites examples including correcting a slipping grasp, locating an object through surface contact, and adjusting motion during insertion when resistance is encountered. Eka's stated goal is to maintain performance as materials and conditions change without engineers retuning the robot for each application. The company frames its approach around three pillars: a single model that generalizes across objects, tasks, and environments; fast and reliable adaptation to unexpected conditions; and safe, collaborative operation alongside humans.
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 deployment on hardware. Li and Raghuram contrasted this with learning from human demonstrations, which they described as time-intensive and poorly suited to conveying the full range of failure modes. The post acknowledges that contact dynamics are difficult to model and that small errors can render a simulated policy ineffective in the real world, identifying the sim-to-real gap as one of robotics' hardest problems.
Eka reports up to 1,500 picks per hour in laboratory induction and bin-picking tasks, roughly two and a half times human throughput, according to the announcement. The company says the system has demonstrated manipulation of fragile objects, slippery industrial parts, and contact-rich insertion tasks. Sustaining these results across full production shifts is identified as the next milestone. Co-founder Pulkit Agrawal, of the Improbable AI Lab at MIT CSAIL, is described as a leading researcher in robot learning, and co-founder Tuomas Haarnoja co-created Soft Actor-Critic at Berkeley and later worked at Google DeepMind on simulation-trained humanoid behaviors. Eka says its broader team includes researchers from MIT, Berkeley, Harvard, CMU, BU, UPenn, DeepMind, Microsoft, and Boston Dynamics. The announcement did not disclose the Series A round size or valuation.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.