Despite investing billions of dollars and feeding countless videos into deep learning models, robots are still not particularly adept at grasping objects.

This situation may be about to change. Eka Robotics has introduced an operational architecture that combines force-sensitive grippers with a new type of deep learning model. The company was founded by Professor Pulkit Agrawal from MIT and former Google DeepMind researcher Tuomas Haarnoja. They have just completed a $150 million Series A financing round, which raised the company's valuation to $1.45 billion. This financing round was led by venture capital firms a16z and F-Prime.

"They have already surpassed Superman-level performance in certain tasks, so we believe that in the coming months, many tasks that require dexterity will also achieve Superman-level performance," Raghu Raguram, a general partner who participated in this round of financing and co-led the investment, told TechCrunch.

In an era where robot demonstrations are emerging in abundance, the capabilities of Eka have shocked seasoned professionals in the robotics field.

A professor at the University of Illinois at Urbana-Champaign, Kris Hauser, stated that compared to other robot demonstrations he has recently seen, the robot Eka is “quite stunning.” Hauser said that this Eka robot “is able to adapt to a large number of different objects without retraining,” and added that it also possesses a kind of “humanization, even a somewhat eerie kind of humanization – if it fails the first time, it will try again and again.”

Another robotics researcher, who requested to remain anonymous due to his work in a cutting-edge laboratory, told TechCrunch that witnessing the actual performance of the Eka robot was an emotional experience, similar to the feeling people had when robots first achieved reliable bipedal walking five to ten years ago. You can check it out for yourself:

What is most noteworthy here is the smoothness and speed of the robot's movements. Eka states that this robot can grasp approximately 1,500 objects per hour, each weighing up to 1 kilogram, while humans can only complete about 600 such grasps; Eka also mentions that the success rate of its grasping actions consistently exceeds 95%. This robotic arm moves so quickly that it has the potential to injure itself. Observers from the aforementioned cutting-edge laboratory commented on this, saying, “This shows that they have great confidence in their algorithms not to damage the robots. I wonder how many robots they have damaged along the way?”

When the researchers from this cutting-edge laboratory saw the robot Eka, they handed it a raspberry – which seemed to be the laboratory’s favorite test item. “A large and heavy robot rushed towards my arm, then slowed down at the last moment and gently touched me. I still feel that sensation in my body to this day,” they told TechCrunch.

How to avoid crushing raspberries

These capabilities stem from the way its underlying model is constructed. Unlike visual-language-action models, which train the relationship between language, visual input, and action; Eka adopts a visual-force-action model, focusing on the relationship between language, force input, and action. Critics of current generative AI robot models, such as Rodney Brooks, argue that such models will fail if they do not attempt to represent force and touch; the results of Eka seem to indicate that these critics have some merit.

"Vision is not the main method to solve the last millimeter of problems," Eka, co-founder with Agrawal, told TechCrunch. "Vision can only roughly tell you where an object is and what its posture is, but the actual grasping, the real searching, is accomplished through force."

Eka had to design a unique robotic hand that is more flexible than traditional grippers and is equipped with more force sensors; it responds to contact, which is also why it doesn't crush the raspberries. For obvious competitive reasons, the founders remain tight-lipped about how they achieved this, but the core of the approach is to use physical simulators to model the various forces that robots encounter when grasping objects, and to create an reinforcement learning environment that can operate on a large scale based on this. They claim that they can generate thousands of hours of robot experience in just a few minutes.

"Rather than being limited to the data collected from human demonstrations and other sources, Eka is advancing this method through reinforcement learning, enabling robots to surpass any demonstration given to them," said Hauser. "What's truly exciting is that they can be unrestricted by the data presented to them for training."

Agrawal and Haarnoja met while pursuing their postgraduate studies at the University of California, Berkeley, and today they are both important researchers in the field of robot control and reinforcement learning. The key bet for Eka is that this technology can be scaled up through simulation, rather than relying on videos taken by gig workers or robot operators over several hours. Simulation is crucial for achieving reliable operation of robots; in the past, this has always been more difficult to achieve for tasks that require interaction with objects, rather than just avoiding them.

“Eka is bridging the gap between universality and performance, a problem that plagues many companies in the modern robotics field,” said Agrawal.

Make general-purpose robots truly useful

Currently, the most generalizable physical AI models – that is, models capable of attempting many different tasks – do not perform particularly well on these tasks. However, if the AI models only excel in a single task, it may not necessarily constitute a viable commercial loop for the company. The promise of generative AI is that robots can quickly learn a variety of tasks, even entirely new ones.

Current methods have already enabled robots to possess more general intelligence than before – for example, the latest large language model Astra can operate robots in unfamiliar environments and complete tasks, but the process is slow and costly. The vision of models like Eka is to handle physical actions, while more general-purpose models are responsible for deciding how to solve problems. For instance, Astra may know that the dishes in the dishwasher need to be placed in the cabinet after they are washed, but a model like Eka would be able to pick them up without breaking them.

“Most people in this field are focused on planning,” said Eka, co-founder of Haarnoja, “so they approach it from top to bottom: first creating high-level intelligence, and then pushing it down to the physical level. But we think from the bottom up.”

However, Eka still needs to prove itself in real-world scenarios. The company is currently conducting secret pilot projects with several industrial clients, involving tasks such as grasping and placement, as well as cable handling. Nevertheless, the transition from the laboratory to the production line is almost as significant as the gap between simulation and reality.

"The real challenges will arise when they collaborate with integrators and move onto actual production lines, where the system needs to run continuously for a year without any failures," said Hauser. Another unresolved question is, "Just how generalizable are these technologies to different tasks?" What he mainly observed were pick-and-place operations, but in many industrial tasks, the truly valuable parts far exceed just picking and placing.

Agrawal also focuses on deployability; he stated that the current gripper hardware cost of Eka is less than $1,000, with a rated lifespan of 5 million working cycles, but the ultimate goal is to reach the 20 million cycle threshold set by standard industrial robotic arms.

"What we're working on is how to design a complete set of work processes as an engineering project, but at the beginning we'll be very practical. Let's focus on these two specific scenarios first: data scraping, organization, placement, and cable management – making them successful cases first," said Agrawal. "Once we succeed in these areas, there's no limit to what we can achieve in the future based on the methods we're developing."