Despite billions of dollars of investment and endless videos fed into deep learning models, robots still aren’t all that good at recognizing things.
That may be about to change, thanks to a manipulation architecture built by Eka Robotics that includes force-sensing grippers and a new kind of deep learning model. The startup, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, just closed a $150 million Series A round led by venture funds a16z and F-Prime, giving it a post-money value of $1.45 billion.
“We’ve already surpassed superhuman performance in certain tasks, so we think we’ll see superhuman performance in many dexterous operational tasks in the coming months,” a16z general partner Raghu Raghuram, who co-led the funding round, told TechCrunch.
In an age of endless robot demonstrations, Eka’s abilities are shocking even veteran robot builders.
Chris Hauser, a professor at the University of Illinois at Urbana-Champaign, said it had a “wow factor” compared to other robot demonstrations he’s seen recently. The Eka robot “was able to adapt to a variety of objects without retraining,” Hauser said, adding that it also had a “humanity, an eerie humanness to it that tries again and again even if it fails once.”
Another roboticist, who requested anonymity because he works in a remote lab, told TechCrunch that seeing Eka’s robots in action was an emotional experience, comparable to when robots first achieved reliable leg locomotion five to 10 years ago. See for yourself:
What should be noted here? How smoothly and quickly the robot moves. They can pick up about 1,500 objects weighing up to 1 kg per hour, compared to about 600 for humans. Eka says it reliably picks up objects over 95% of the time. And the robot arm moves so fast that it has enough strength to hurt itself. This, according to an observer at Frontier Labs, “puts a lot of faith in an algorithm that doesn’t break robots. How many robots did they break along the way?”
When a Frontier Lab roboticist encountered Eka’s robot, he handed it a raspberry, the lab’s favorite test item. “The experience of a large, heavy robot rushing towards my arm, slowing down at just the right moment to gently touch me, is like having this feeling inside my body while it speaks,” they told TechCrunch.
How not to crush raspberries
These features are a result of how the basic model is constructed. The schema here is a visual-force-behavior model that focuses on the relationship between language, force input, and movement, rather than a visual-verbal-behavior model trained to understand the relationship between language, visual input, and movement. Critics of the current direction of generative AI robotics models, such as Rodney Brooks, say they will fail if they don’t try to express force or touch. Eka’s results suggest that such critics have a point.
“Visibility is not the primary way to solve this last millimeter problem,” Eka co-founder Agrawal told TechCrunch. “Vision tells us roughly where objects are and what pose they are in, but the actual grasping and searching is done through forces.”
Eka had to design a unique robotic hand that was more flexible and had more force sensors than traditional grippers. It doesn’t crush the raspberries because it responds to contact. The founders are tight-lipped about how they achieved it for obvious competitive reasons, but the bottom line is that they use a physics simulator to model the multiple forces a robot encounters when trying to grab something, and then use that to create a reinforcement learning environment that can be run at scale. They say they can generate thousands of hours of robotic experience in minutes.
“Instead of being limited to data collected from things like human demonstrations, Eka is advancing this approach with reinforcement learning that allows robots to go beyond any kind of demonstration that is given to them,” Hauser said. “What’s really interesting is that you’re not limited to the data presented to you for training.”
Agrawal and Harnoja, who first met in graduate school at the University of California, Berkeley, are leading researchers in the field of reinforcement learning for robot control. Eka’s main bet was that the technology could be run at scale through simulation, rather than relying on video collected by gig workers or robot operators performing tasks over long periods of time. Simulation has been key to achieving reliable operation of robots, but until now simulation has proven more difficult for tasks that require interacting with objects rather than avoiding them.
“What EKA bridges is the versatility and performance gap that plagues much of modern robotics,” Agrawal asserts.
Make general-purpose robots convenient
Currently, the most generalizable physical AI models (models that can try a variety of tasks) are not very good at it. However, an AI model trained to be highly capable at a single task cannot solve an enterprise’s business case. The promise of generative AI is that robots can quickly learn multiple tasks, even new ones.
Current approaches are generally better at creating intelligent robots than before. ChatGPT’s newest LLM, Astra, allows you to operate robots to perform tasks in unfamiliar environments, but it is time consuming and costly. The promise of models like Eka is that they handle physicality while more general models make decisions about how to solve problems. For example, Astra may know that dishes pile up in the cupboard when they come out of the dishwasher, but Eka’s model can pick them up without breaking them.
Haarnoja, co-founder of Eka, says, “Most of this field is focused on planning, so they approach from the top, which means pushing high-level intelligence down to the physical. Whereas we think from the bottom up.”
However, Eka still needs to prove herself on the field. The company is secretly piloting it with a number of industrial customers that do pick-and-place and wire operations, but the leap from lab to line is almost as large as the gap between simulation and reality.
“The big challenge is working with an integrator on the actual assembly line, which has to run uninterrupted for a year,” Hauser said. Another open question is, “To what extent are these techniques generalizable to different tasks?” What I’ve seen is pick and place, but many value-add industry tasks involve more than just pick and place. ”
Agrawal also focuses on deployability. Currently, Eka’s gripper hardware (priced at less than $1,000) is rated for 5 million duty cycles, but he says the ultimate goal is to reach the 20 million cycle benchmark set on standard industrial arms.
“What we’re building is the engineering of how to design the entire workflow, but we’re starting with something very practical, and there are two areas: pick, place, place and wire handling, so let’s make those home runs,” Agrawal said. “Once we hit a home run with them, there’s no limit to what we can do with the approach we’re developing.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
