About 1,500 objects an hour, each weighing up to 1 kg, is the pickup rate for Eka Robotics’ arm. A human manages about 600. Eka reports a success rate above 95%, using a manipulation system that pairs a force-sensitive gripper with a new type of deep-learning model.
MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja founded the company, which has now raised $150 million in a Series A round. Led by venture funds a16z and F-Prime, the financing gives Eka a post-money valuation of $1.45 billion.
“They are already exceeding superhuman performance for certain tasks, and so we think over the next few months there will be superhuman performance in many dexterous manipulation tasks,” Raghu Raguram, the a16z general partner who co-led the fundraising round, told TechCrunch.
Billions of dollars of investment and large amounts of video training haven’t made picking things up easy for robots. But Eka’s demonstration stood out to University of Illinois Urbana-Champaign professor Kris Hauser. He described its “wow factor” relative to the other robot demonstrations he’d seen recently.
Hauser said the robot “was able to adapt to a large number of different objects without retraining,” and noted how it kept attempting a task after failing. He described that quality as “humanness, creepy humanness that it would try over and over again if it failed the first time.”
A second roboticist requested anonymity because they work at a frontier lab. They told TechCrunch that the system prompted an emotional reaction comparable to watching robots first master reliable legged locomotion five to ten years before. The arm’s movement is fast and smooth, and its power is sufficient to damage it at those speeds. That led the observer to conclude that “they have a lot of trust in their algorithm not breaking the robot. I wonder how many robots they broke along the way?”
When the anonymous roboticist handed the robot a raspberry, apparently one of the lab’s preferred test objects, the arm rushed toward their arm, slowed just in time, and touched them gently.
“The experience of a big, heavy robot rushing towards my arm and then slowing down just in time and very gently touching me is a feeling that I am, as we speak, carrying this feeling in my body,” they told TechCrunch.
How to keep a raspberry intact
That precision stems in part from the model’s underlying design. Vision-language-action models train on the relationships among language, visual inputs, and actions. Eka’s alternative is a vision-force-action model, described by TechCrunch as connecting language, force inputs, and actions. Rodney Brooks and other critics contend that generative AI models for robotics won’t succeed without representing touch and force. Eka’s performance lends weight to their argument.
“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.”
Compared with conventional grippers, the hand Eka developed is more flexible and has additional force sensors. It can react to contact, allowing it to hold a raspberry without squashing it. For competitive reasons, the founders haven’t disclosed the implementation in detail, but they have explained the broad training method.
A physics simulator models the various forces a robot encounters when grasping an object. This becomes a reinforcement learning environment that can run at scale. Eka says a few minutes of simulation can generate thousands of hours’ worth of robot experience.
The founders first crossed paths during graduate school at the University of California, Berkeley. Agrawal and Haarnoja are both leading researchers working on reinforcement learning for robotic control. They bet on using simulation to scale this learning, instead of relying on videos collected by gigworkers or robot operators carrying out tasks over many hours.
For Hauser, the advantage is that learning needn’t stop at what the examples show. “Rather than being limited to the data that you’re collecting from, say, human demonstrations, Eka is pushing through reinforcement learning this approach where the robot can surpass any kind of demonstrations that are given to it,” he said. “What’s really exciting is that they’re able to not be limited to the data that is presented to them for training.”
Eka’s dependable object handling relied on simulation. Historically, transferring simulated skills to the physical world has been harder when a task requires contact with objects, rather than avoiding them.
“What EKA is bridging is the generality-performance gap, which plagues a lot of modern robotics,” Agrawal argues.
Putting a general-purpose robot to work
A model’s ability to try a wide range of tasks doesn’t guarantee good results. The most generalizable physical AI models today have trouble carrying tasks out. At the other extreme, models trained to be highly capable at a single task may not close the business case for a company. The promise of generative AI is that robots could quickly learn multiple tasks, even novel ones.
Robots have become more generally intelligent through existing methods. As an example, TechCrunch points to Astra, which its report describes as ChatGPT’s latest LLM. It can use a robot in an unfamiliar environment to carry out tasks, though execution is slow and expensive. The proposed division of labor gives Eka responsibility for the physical actions and a more general model responsibility for deciding how to solve the problem.
Astra, for instance, might know to put plates in a cupboard after they’re taken out of a dishwasher. Handling a dish gently enough to lift it without damage would fall to Eka’s model.
“Most of the field has been focusing on the planning,” Haarnoja says, “so they approach from the top: high-level intelligence, pushing that down towards the physical. Whereas we think about it from the bottom up.”
Eka hasn’t yet proved itself in industrial deployment. It is running secretive pilots with several industrial customers on pick-and-place and wire manipulation. The distance between laboratory performance and production-line use is nearly as great as the one between simulation and the real world.
“What’s going to be a big challenge is when they work with integrators on real assembly lines, and it needs to run without breaking for an entire year,” Hauser said.
How generalizable Eka’s techniques are to different tasks is another question Hauser considers unanswered. “What I’ve seen is pick and place, but many industry tasks that are value added involve more than just pick-and-place.”
Agrawal’s work on deployment also includes hardware durability. According to him, the gripper costs under $1,000 and is currently rated for 5 million duty cycles. The ultimate goal for the gripper hardware is 20 million cycles, matching the benchmark set by standard industrial arms.
“What we are building is that engineering of how to design entire workflows, but we are starting off being very pragmatic, and saying here are two verticals in pick, arrange, place, and wire handling, let’s make them be home runs,” Agrawal said. “Once we have made home runs on them, the sky is the limit in how much we can do with the approach we are developing.”
