Eka Robotics unveils VFA dexterity model

Robot dexterity is the annoying last problem in automation. Getting a machine to move is easy enough. Getting a machine to touch the world quickly, adapt mid-motion, and not crush the thing in its hand is much harder. That is the gap Eka Robotics says it can finally narrow. On April 29, the Cambridge startup came out of stealth and unveiled VFA — short for vision-force-action — with demos that looked unusually fluid for real-world manipulation. (ekarobotics.com) ### What did Eka actually launch? Eka launched itself, basically. The company had been quiet, then used the April 29 reveal to introduce both the startup and its core model family, VFA. The founders are Pulkit Agrawal, an MIT EECS associate professor, and Tuomas Haarnoja, a former DeepMind researcher. Eka says the goal is broad physical intelligence for robots, but the first wedge is dexter(ekarobotics.com) factories, kitchens, and warehouses hard to automate. (people.csail.mit.edu) ### What is a vision-force-action model? Most people have now heard of vision-language-action robots — systems that look at the world, parse a text instruction, then act. Eka is pushing a different stack. Its claim is that for manipulation, force is the more important signal than language. A robot does not just need to know that an object is a bulb or a berry. It needs to(people.csail.mit.edu)y. That is what “vision-force-action” is getting at — seeing the scene, sensing interaction forces, and turning both into motion. (humanoidsdaily.com) ### Why is force the hard part? Because the world pushes back. Pick-and-place is easy when every box is rigid, identical, and sitting where you expect. Real work is messier. Objects roll, bend, snag, wobble, and break. The last few millimeters of motion are where most failures happen — when threads(humanoidsdaily.com)e it by slowing way down. Eka’s whole pitch is that slowing down should not be necessary. (humanoidsdaily.com) ### What did the demos show? The demos were chosen to make that point obvious. Eka showed a robot arm screwing in a delicate light bulb, reacting to a rolling bulb on a table, sorting unfamiliar objects with a simple two-finger grasp, and handling food items like chicken nuggets on a moving line. O(humanoidsdaily.com)cision. It was precision at speed. (humanoidsdaily.com) ### How is Eka training this? Not mainly by copying humans. Eka says it leans on high-fidelity simulation, reinforcement learning, and custom tactile hardware, then transfers those policies into the real world. That matters because human video can teach broad task structure, but it is weak at the a(humanoidsdaily.com)r more often — and more dangerously — than a real lab ever could. (humanoidsdaily.com) ### Why does this stand out right now? Because the robotics zeitgeist has tilted toward language. A lot of the excitement lately has gone into generalist systems that can take spoken or text instructions and do many tasks passably well. Eka is arguing that this trend may miss the bottleneck. If you(humanoidsdaily.com) fingertips.” (humanoidsdaily.com) ### So what is the catch? A demo is not deployment. Eka still has to prove that VFA scales across hardware, survives long factory hours, and works on enough tasks to justify a foundation-model label. But the reason people noticed this launch is simple — the motions did not look like the usual caref(humanoidsdaily.com)hat always arrives later. (wired.com)