Most robot videos you scroll past these days blur together: a humanoid in a warehouse, a hand folding a shirt, some jazz on the soundtrack. What as reported by TechCrunch happened at Eka Robotics’ demo was different enough that veteran robot builders got emotional watching it. An arm rushes at a person’s body at full speed, then slows “just in time and very gently” to make contact — feeling where the object is rather than guessing from a camera. One frontier-lab roboticist described the experience as comparable to watching robots first master legged locomotion five to ten years ago.
The Cambridge, Massachusetts startup has just closed a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime. The founders are reinforcement-learning people from the deep end of the field: MIT professor Pulkit Agrawal, who runs the Improbable AI Lab at CSAIL, and Tuomas Haarnoja, who co-created the Soft Actor-Critic algorithm at Berkeley and later trained humanoid behaviors at Google DeepMind.
Force is the native language
Eka’s core argument is that the field’s dominant architecture is built on the wrong sense. Most frontier manipulation models are vision-language-action (VLA) systems: they watch thousands of hours of human video and learn what a task looks like. Eka’s model is vision-force-action (VFA) — vision still tells the robot roughly where an object sits, but the actual grasp, the searching, the “last millimeter” of contact, is negotiated through force feedback.
“Vision is not the primary way in which you close this last millimeter problem,” Agrawal told TechCrunch. “Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces.”
The numbers behind the claim: Eka’s robot picks up roughly 1,500 objects an hour (a human manages about 600), with over 95% reliability on items up to 1 kg, and it adapts to new objects without retraining. The gripper hardware costs less than $1,000 and is rated for 5 million duty cycles — well short of the 20 million-cycle bar set by standard industrial arms, a gap the company is honest about.
The AlphaZero bet
Training data is where Eka diverges even more sharply from the pack. Companies like Generalist AI and Sunday Robotics collect hundreds of thousands of hours of real-world footage via human-worn gloves. Eka trains its manipulation policies almost entirely in physics simulation, with reinforcement learning letting virtual robots practice millions of variations in friction, grip and contact — experience failure modes humans rarely demonstrate, then transfer the learned policy to real hardware. The comparison the company itself draws is AlphaZero: superhuman performance through self-play rather than imitation.
Kris Hauser, a University of Illinois Urbana-Champaign robotics professor who saw the demo, said the system “was able to adapt to a large number of different objects without retraining” — and noted it can “surpass any kind of demonstrations that are given to it,” because the robot is not limited to the data humans can perform for a camera.
Why it matters for the robot-brain market
This raise lands in a market segment that is heating up fast. DeepSeek is reportedly leading a seed round for Winverse AI, an ex-Tencent-scientist robot-lab startup valued near $400 million, while Skild AI’s S1 foundation model and Generalist’s $3 billion round already made “robot brains” an investor category of their own. Eka’s thesis gives that category an organising idea: the hands and the planner deserve different architectures, and force — not language — is the missing modality. It rhymes with what Boston Dynamics did at the hardware end, cutting Atlas’s pinky to re-engineer the hand around simulation-first force control.
Both interpretations deserve airtime. The optimist reading: dexterity has been the field’s hardest problem for decades, and a team combining MIT sensorimotor research with DeepMind RL training lineage is unusually credible to bet on. The cautious one — which TechCrunch’s own reporting supplies — is that Eka’s pilots with industrial customers are secretive, the duty-cycle rating is a fifth of industrial standard, and the leap from lab demo to production line is “almost as big as the gap between sim and real.” A $1.45B valuation is priced on the first reading; the founders’ own pragmatism (“here are two verticals… let’s make them be home runs”) is built on the second.
For anyone tracking where physical AI actually gets deployed, the interesting signal is not the valuation itself. It is that the money is arriving after the capability demos rather than before them — as first shown in Humanoids Daily’s launch coverage of the company’s stealth exit — and after the gripper’s economics were benchmarked against human warehouse workers. That ordering has been rare in the 2026 robotics funding cycle.
Sources: TechCrunch, a16z announcement, Humanoids Daily, Eka Robotics
