Eka, a Cambridge, Massachusetts startup founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, is demonstrating robot hands that screw in light bulbs, sort chicken nuggets and grasp keys on a plush key ring without dropping them. None of the few dozen robot arms commercially available today can screw in a light bulb. Eka's pitch is that dexterity, the last unsolved problem in robotics, is now a scaling question rather than a research one.
The company's office sits a short walk from MIT, packed with robot arms, grippers and tables of test objects ranging from earplug boxes to hairbrushes. In WIRED's demos, the robot nipped at items a few times before lifting them, then resisted briefly when a reporter tried to take a key ring back. "A couple of years ago, we realized that dexterity can finally be cracked," Agrawal said. He frames the stakes bluntly: "Trillions of dollars flow through the human hand. To me, this is the biggest problem in the world to be solved."
The reference point for how hard this is remains OpenAI's Dactyl, unveiled in October 2018, four years before ChatGPT. Dactyl used a Shadow Robot hand and reinforcement learning in simulation to solve a Rubik's Cube, but the cube was custom-built with sensors tracking each face. The robot couldn't recover if the cube slipped, couldn't work from imperfect starting angles, and couldn't handle anything other than that one rigged object. OpenAI shelved its robotics work to focus on large language models, and only recently restarted the effort.
Key facts
- 01Eka was cofounded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, based in Kendall Square, Cambridge.
- 02By late 2021, Agrawal's lab had built a virtual hand capable of manipulating 2,000 objects upside down in simulation.
- 03OpenAI's Dactyl, launched October 2018, could only manipulate a single sensor-equipped Rubik's Cube and was abandoned before ChatGPT shipped four years later.
- 04Eka trains robots through thousands of computer hours in simulated worlds, an approach modeled on Google DeepMind's AlphaZero.
- 05None of the few dozen robot arms on the market today can screw in a light bulb, a task Eka's prototype performs.
Agrawal told WIRED that members of the Dactyl team considered the simulation approach a dead end because of the sim-to-real gap — the chasm between behavior in a virtual environment and behavior with real physics. When he described his own simulation project to a former Dactyl engineer, he says, he got "a one-hour lecture from them saying, 'This will never work.'" He kept going.
“By late 2021, Agrawal had built a virtual hand capable of manipulating 2,000 objects upside down — eight years after OpenAI's Dactyl could only handle a sensor-rigged Rubik's Cube.”— Jaeden Schafer
By late 2021, Agrawal's lab had built a virtual hand capable of manipulating 2,000 objects upside down. Haarnoja, working separately at Google DeepMind, was using virtual reinforcement learning to train small humanoid robots to play soccer — arguably harder than light-bulb manipulation, since the field doesn't roll under the players' feet. The two had originally met as graduate students at UC Berkeley and teamed up to start Eka.
Their approach is the contrarian one in robotics right now. The well-funded camp, including most of the humanoid startups, is training vision-language-action models on huge volumes of video showing humans folding T-shirts or teleoperating robots. A small industry has emerged paying people to wear motion-capture gloves and cameras while doing routine tasks. Eka skips that entirely. Its robots spend thousands of computer hours practicing inside simulated worlds and inventing their own solutions.
Agrawal compares the philosophy to AlphaZero, the Google DeepMind system that learned chess and Go from self-play and discovered new strategies along the way. Eka has built custom grippers that incorporate a sense of touch and a new algorithm it calls a vision-force-action model — a simulator that includes mass, inertia and the way an object's weight interacts with the robot's grip, not just how the pixels move on screen. The founders won't disclose details, citing commercial sensitivity.
Ken Goldberg, the UC Berkeley robotics professor advising Eka, has known Agrawal since his graduate-student days. "Pulkit is a very creative thinker," Goldberg said. "He's always pushing in a direction that other people aren't." The two first crossed paths at a 2017 AI conference in Long Beach, California, where Agrawal had just published a paper on computers learning to play video games.
The Austrian computer scientist Hans Moravec observed in the late 1980s that tasks humans find trivial — picking up a key ring, screwing in a bulb — are the hardest things to teach a machine, while abstract reasoning is comparatively easy. That observation, known as Moravec's paradox, has held up for nearly four decades. The fastest humans solve a Rubik's Cube in about three seconds; in those same three seconds, a computer can solve thousands of variations of the puzzle in software. Bridging that gap into the physical world is what Eka is selling.
The skeptical case is straightforward: every robotics startup of the last decade has shown demos that look generally intelligent and then struggled to generalize beyond the demo table. Dactyl looked superhuman in the press release and turned out to be brittle. Eka's founders won't disclose how they close the sim-to-real gap, which makes their claims hard to evaluate from outside. Vision-language-action competitors with more funding and more real-world data may simply outscale them.
If Eka's simulation-first approach works, it changes the economics of robotics in a specific way. The VLA camp needs ever-larger fleets of motion-capture-rigged humans to generate training data, an expensive bottleneck that scales linearly with capability. A self-play system that learns inside a physics simulator scales with compute, the resource AI companies already know how to buy in bulk. Agrawal's framing — "Some people want robots to be human-level. For us the goal is superhuman" — only makes sense if the underlying method has the same compute-driven trajectory that took language models from novelty to 1.8M paying users on a single product. Whether Eka's pincers actually have that trajectory is the question its next round of demos will have to answer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
