How many robots does it take to screw in a light bulb? If we stick to the robotic arms currently on the market, the answer is zero. Until now, robotic appendages lacked the dexterity to carry out complex and delicate actions. Especially when handling delicate materials. Such limitations have been a boon to advocates of manual labor who hoped to keep large-scale automation at bay. However, a new robotic arm from startup Eka could provide researchers with a solution to one of robotics’ most intransigent questions and pave the way for untold scales of industrial automation.
Based in Cambridge, Massachusetts, Eka offers a unique approach to the problem of dexterity in robotics. Founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, the company takes a unique approach to teaching its AI systems to perform complex movements. By deploying what its leaders call a Vision Force Action, or VFA, learning model, Agrawal says Eka’s AI system could create “a new foundation bringing together performance, generality and safety to put high-performance robots in everyone’s hands.”
A recent report from Wired’s Will Knight attests to this potential. By showing how Eka’s robotic arm performs complex tasks, from manipulating keys to sorting chicken nuggets, the report could prove prophetic for a rapidly changing industrial landscape. AI-based robotics is already revolutionizing manual labor, particularly in China, where robots often perform tasks alongside workers in factory farms and factories. Some advances have given rise to dark factories, where products such as smartphones can be manufactured around the clock, without human workers. Despite these changes, however, skillful tasks remain out of reach. Unfortunately for global labor markets, a robot capable of handling raspberries could be a game-changer.
Tackling the dexterity problem
Developers typically train autonomous robots using vision-language-action (VLA) models. Similar to training large language models, VLAs feed AI systems massive amounts of visual information until they learn to reproduce it. But this approach has obvious drawbacks. For example, what happens if something unexpected happens?
To illustrate this, take a piece of fruit and think about the infinite variations that an action could cause. Strawberries, for example, come in different shapes, sizes and levels of ripeness. They may be wet, placed at odd angles, or mushy. The natural laws of force, gravity and inertia amplify these problems. How does a robotic hand fit into a malformed berry? What happens if he slips or falls? The answer is more data. But to get it, researchers must tackle the gargantuan task of feeding their models videos showing the countless ways in which a berry is picked. Thus, the problem of robot dexterity is rooted in the most fundamental fact of our existence: the infinite variations of our natural world.
However, Eka takes a different approach. To train its robotic appendage, the startup uses a vision force action model, in which an AI model runs thousands of hours of simulations that force the system to solve these problems directly. Eka’s VFA simulations include natural principles such as mass and inertia to teach the robotic system how these forces act on the objects it manipulates. In his report, Knight compared such a system to those powering Google’s Alpha Zero program, which has discovered new chess strategies. Unfortunately, the duo remains strategically tight-lipped about how they translate these simulations into the real world, an obstacle known as the simulation-reality gap in robotics. However, the proof is in place. Or, in Eka’s case, chicken nuggets.
Dexterity or misfortune?
To test Eka’s robotic hands, Knight presented her with a series of objects to collect, including earplugs, a hairbrush, and keys. Although Knight acknowledges that the robot often takes “a few pinches” to accomplish its task, the hand ultimately accomplished its task thanks to a set of custom pincers that give the robot feel. The company’s videos document Knight’s reporting, showing the robot screwing in light bulbs and retrieving a strawberry from a model’s open palm.
In one example, Knight watched Eka sort a table of chicken nuggets scattered into a conveyor belt of moving containers, showcasing the improvisation skills that are typically lacking in industrial robotics. Although banal, these examples are significant. As Knight points out, food handling is largely done by human operators due to its delicate composition. Incredibly, chicken nuggets could be the precursor to mass industrial automation. According to Wired, Eka thinks it’s “halfway there.” Now it’s a question of scale.
Inevitably, such advances must be framed in the context of massive job losses. According to a 2026 Goldman Sachs report, approximately 300 million jobs are exposed to AI automation. Meanwhile, the World Economic Forum reported that 58% of employers believe advances in robotics and automation will transform their businesses by 2030. And although Eka touts its robots as performing “mastery alongside humans,” its founders’ language attests to the technology’s harmful economic consequences. For example, Agrawal told Wired that the “biggest problem in the world to solve” was that “trillions of dollars flow through human hands.” While the professor probably didn’t mean that literally, his technology undeniably has the potential to wrest that money right from workers’ fingertips. Eka has yet to indicate whether his next major “problem” will be the next claw-sized hole in workers’ wallets.
