Robotic grasping has long been a stubborn problem in automation—despite vast research spending and enormous volumes of training video, machines still lag far behind human dexterity when it comes to picking things up. Eka Robotics, backed by a fresh $150 million Series A led by a16z and F-Prime, may have just cracked it.
The startup, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, now sits at a $1.45 billion valuation post-money. Its breakthrough centers on a custom force-sensitive gripper paired with what the team calls a vision-force-action model—a departure from the vision-language-action approach that dominates current AI robotics.
Why Touch Matters More Than Sight#
The conventional wisdom in AI robotics emphasizes visual understanding: what the robot sees determines what it does. Eka's founders argue that this misses the critical final step. According to Agrawal, "Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces."
This insight aligns with longstanding critiques by robotics pioneers—including critics of vision-first approaches—who have warned that touch and force feedback are essential to solving precision manipulation. By prioritizing tactile feedback alongside visual input, Eka's system can perform feats like gently grasping a raspberry without crushing it, something that demands millimeter-level control.
Simulation-Powered Learning at Scale#
The technical achievement rests on using a physics simulator to model the forces a robot encounters during grasping, then running reinforcement learning in that simulated environment. This sidesteps a major bottleneck: the need to collect endless hours of human demonstrations or data from costly real-world robot trials. Instead, Eka claims it can generate thousands of hours of robot experience in minutes.
The strategy represents what Agrawal frames as a "bottom-up" approach. Rather than building high-level decision-making and pushing it down to physical execution, Eka perfects the physics layer first, leaving space for general-purpose AI systems to decide what to do while Eka's model handles how to do it reliably.
The Numbers#
In tests, Eka's robot picks up roughly 1,500 objects weighing up to 1 kilogram per hour—compared to about 600 for a human operator. It succeeds more than 95% of the time and moves with such speed and confidence that it could injure itself if its control algorithms failed. That confidence reflects deep trust in the underlying model.
Independent roboticists have voiced enthusiasm. Kris Hauser from the University of Illinois Urbana-Champaign described a "wow factor" in Eka's demonstrations, noting the robot's ability to adapt to diverse objects without retraining and its persistence when tasks fail. Another prominent roboticist, speaking anonymously due to institutional constraints, compared watching Eka's robot to the emotional milestone when legged locomotion became reliable five to ten years ago.
The Road Ahead#
Eka faces the classic lab-to-industry gap. The company is piloting its system with select industrial partners on pick-and-place and wire-handling tasks, but scaling to production lines—where robots must run reliably for a year without downtime—presents a different engineering challenge. The startup's gripper hardware, priced under $1,000, currently supports 5 million duty cycles; the target is 20 million to match standard industrial arms.
Hauser also flagged an open question: how well these techniques generalize beyond pick-and-place to the varied, value-added tasks that dominate real assembly lines. Agrawal has indicated pragmatism about sequencing, focusing first on two verticals—object arrangement and placement, plus wire handling—before expanding scope. "Once we have made home runs on them," he said, "the sky is the limit in how much we can do with the approach we are developing."
Learn more: Read the full TechCrunch story