Eka is worth $1.45B thanks to a physical AI model that has experts raving

Eka Robotics closed a $150 million Series A funding round, giving the company a post-money valuation of $1.45 billion. The round was led by venture funds a16z and F-Prime. Post-money means the valuation includes the new investment.

The funding supports Eka’s manipulation architecture, which combines a force-sensitive gripper with a deep-learning model. Its founders are MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja. Their system focuses on helping robots grasp and handle objects reliably.

The investment reflects strong confidence in Eka’s progress. The article reports that the robot can handle about 1,500 objects per hour and succeeds more than 95% of the time. An a16z general partner predicted superhuman performance across many dexterous manipulation tasks within months, though the article does not describe how Eka will spend the funding.

Physical AI refers to artificial intelligence that interacts with the physical world through sensors, machines, and actions. It must respond to real objects, contact, movement, and changing conditions. That makes it different from AI that only generates or analyzes text, images, or video.

Eka’s robot shows why this distinction matters. It must locate an object, move toward it, feel contact, adjust its grip, and lift it without causing damage. Vision helps estimate an object’s position and pose, but force inputs help solve the final part of the grasp.

The article presents physical AI through Eka’s vision-force-action model. The system connects visual information and force with robot actions. This approach could make robots more useful in tasks involving delicate or varied objects. Eka’s reported ability to adapt to many objects without retraining suggests progress, although the article describes demonstrations rather than universal performance.

Eka’s robot can pick up about 1,500 objects in one hour, provided they weigh up to 1 kilogram. Eka says it succeeds reliably more than 95% of the time. Those figures highlight both speed and consistency in a task that has challenged robots.

The comparison in the article is striking. A human can pick up about 600 objects in an hour, according to the report. Eka’s robot therefore handles the stated test workload at more than twice that rate. Its movements are also described as smooth and rapid, even though the arm is strong enough to hurt itself.

These numbers show a promising demonstration, not proof that every robotic task is solved. The article says Eka adapted to many different objects without retraining. It also reports that observers were surprised by the robot’s humanlike persistence after failures. Eka argues that its method can improve generality and performance together.

Delicate objects can be damaged when a robot applies too much force. Vision can estimate where an object is and how it is positioned, but it cannot fully determine the pressure needed during the final contact. Force and touch sensing provide information at that critical moment.

Eka designed a more flexible robotic hand with additional force sensors. When the hand contacts an object, the sensors help the robot respond to the forces it encounters. In the article, a frontier-lab roboticist handed the robot a raspberry. The robot rushed toward the person’s arm, slowed in time, and gently touched the person.

Eka trains this behavior using a physics simulator that models multiple forces during grasping. Reinforcement learning then lets the robot practice responses in a simulated environment. The result is a system designed to search, close, and adjust through force, rather than simply moving to a visually estimated location.

Many robotics systems use vision-language-action models. These models learn relationships among language, visual inputs, and actions. Eka instead uses a vision-force-action model, which emphasizes how visual information and physical forces should guide movement.

The difference matters during grasping. Vision can tell the robot roughly where an object is and what pose it has. But the robot still needs to search for contact, feel resistance, and adjust its grip. Eka co-founder Pulkit Agrawal said those final actions happen through forces rather than vision alone.

Eka’s approach combines a force-sensitive gripper with deep learning and simulated practice. Its hand reacts to contact, which helps explain the raspberry demonstration and its reported success with varied objects. The article presents this as an alternative to relying mainly on videos or language-linked demonstrations. It does not claim that vision-language-action models are useless, only that force is important for dexterous manipulation.

Human demonstrations are limited by the time people spend collecting them. Eka instead models grasping in a physics simulator, including the multiple forces a robot may encounter while handling objects. This creates a practice environment that can be run at large scale.

Reinforcement learning lets the simulated robot try actions, experience outcomes, and improve its behavior. Eka says it can generate thousands of hours of robot experience in minutes. The system can therefore explore many grasping situations without requiring human operators to perform every attempt.

The approach also lets the robot go beyond copying demonstrations. Kris Hauser said Eka was not limited to the data presented during training and could surpass demonstrations through reinforcement learning. The article identifies simulation as key to reliable manipulation, especially when robots must interact with objects. It does not provide the simulator’s detailed design, because Eka’s founders are keeping parts of it confidential.

A general-purpose robot must handle many different objects and situations. A highly reliable robot must also respond precisely when contact, shape, weight, or resistance changes. The article describes this tension as the generality-performance gap, which affects many modern robotics systems.

Eka’s robot reportedly adapted to a large number of objects without retraining. Its flexible, force-sensing hand responds to contact, while the vision-force-action model connects perception and physical feedback to movement. Physics simulation exposes the robot to many grasping forces, and reinforcement learning lets it improve through large-scale practice.

This combination could help robots become both broader and more dependable. Eka says its robot picks up about 1,500 objects per hour with more than 95% reliability. Observers were impressed, but the article presents these as current capabilities rather than a finished solution. Eka’s founders expect the approach to support superhuman performance in more dexterous tasks.