Eka Robotics Secures $1.45B Valuation for Physical AI Manipulation Models

Despite massive investment in deep learning, robots have historically struggled with the delicate task of picking up objects.

A manipulation architecture developed by Eka Robotics, featuring a force-sensitive gripper and a novel deep-learning model, aims to resolve this issue.

Founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, the startup recently closed an $150 million Series A funding round, leading to a $1.45 billion post-money valuation. The round was co-led by venture capital funds a16z and F-Prime.

Raghu Raguram, an a16z general partner and co-leader of the fundraising, expressed confidence in the technology, stating, "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."

In a landscape saturated with robotic demos, Eka's capabilities have left veteran robot builders astonished.

Kris Hauser, a professor at the University of Illinois Urbana-Champaign, noted a "wow factor" compared to other recent demos. He highlighted that the Eka robot could adapt to various objects without retraining, while also exhibiting a "humanness, creepy humanness that it would try over and over again if it failed the first time."

An anonymous roboticist from a frontier lab described the experience of seeing Eka's robot in action as emotional, comparable to the breakthroughs in reliable legged locomotion seen a decade ago.

The standout feature is the robot's smooth and rapid movement. It can handle approximately 1,500 objects weighing up to 1 kg per hour, surpassing the human capability of about 600 objects. Eka claims a reliability rate of over 95%, though the robot's speed is so high it can occasionally hurt itself during movement.

Reflecting on a specific interaction where the robot approached his arm with a raspberry, the anonymous roboticist described the sensation as "a feeling that I am, as we speak, carrying this feeling in my body."

This performance stems from a unique foundational model that utilizes a vision-force-action schema rather than the typical vision-language-action model. This distinction addresses concerns raised by critics like Rodney Brooks, who argued that generative AI robotics models fail without a representation of force and touch.

"Vision is not the primary way in which you close this last millimeter problem," Eka co-founder Pulkit Agrawal explained. "Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces."

To achieve this, Eka designed a flexible robotic hand with enhanced force sensors capable of reacting to contact to prevent crushing objects. The company utilizes physics simulators and scaleable reinforcement learning environments to generate extensive training data, although specific details regarding the hardware design remain guarded.

Tuomas Haarnoja, another co-founder, emphasized the shift away from relying solely on human demonstrations. "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."

Agrawal and Haarnoja, former graduate school peers at UC Berkeley, are experts in reinforcement learning for robotics control. Their strategy relies on simulation rather than gigworker videos, aiming to solve the "generality-performance gap" that plagues modern robotics.

While current general models like Astra are slow and costly, Agrawal argues that Eka's model handles physicality while a general model handles decision-making. "Most of the field has been focusing on the planning... whereas we think about it from the bottom up," Haarnoja added.

Despite the lab success, Eka faces challenges in real-world deployment. The company is currently conducting secret pilots with industrial clients for pick-and-place tasks, but bridging the gap between simulation and a continuous, unbroken year of operation on assembly lines remains a significant hurdle.

Hauser points out that industry tasks often involve more than just pick-and-place. "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."

Agrawal is pragmatic about immediate deployment, focusing on two verticals: pick, arrange, place, and wire handling. He aims to achieve "home runs" in these areas before expanding the approach to more complex tasks.