Eka Robotics, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, closed a $150 million Series A led by a16z and F-Prime, valuing it at $1.45 billion post-money. The startup has built a manipulation architecture that pairs a force-sensitive gripper with a new kind of deep-learning model. a16z general partner Raghu Raghuram said the system already exceeds superhuman performance on certain tasks. He expects superhuman performance across many dexterous manipulation tasks in the coming months.
University of Illinois Urbana-Champaign professor Kris Hauser said the robot adapted to a large number of different objects without retraining, and it kept retrying when it failed. An anonymous roboticist from a frontier lab compared seeing it to the moment robots first achieved reliable legged locomotion. The robot picks about 1,500 objects weighing up to 1 kg per hour, compared with about 600 for a human, and Eka says its pick success rate is above 95%. Its arm moves fast enough to damage itself, which suggests strong confidence in the control algorithm.
Eka did not build a vision-language-action model. Its foundation model is a vision-force-action model centred on force inputs and actions. Agrawal said vision only tells the robot roughly where an object is and how it is positioned, while the actual grabbing in the last millimetre happens through forces. Eka designed a more flexible hand with more force sensors that reacts to contact, which is how it avoids crushing a raspberry. Eka's results lend support to critics such as Rodney Brooks, who argue that generative AI robotics will fail unless it represents force and touch.
Eka uses a physics simulator to model the forces a robot meets while grasping, and turns it into a reinforcement learning environment that runs at scale. The company says it can generate thousands of hours of robot experience in minutes. Hauser noted that this lets the robot go beyond any human demonstrations given to it, rather than being limited to collected training data. Agrawal says the approach bridges the generality-performance gap that plagues much of modern robotics.
Today the most general physical AI models can attempt many tasks but don't do them well, while single-task models may not make the business case work. A general model like ChatGPT's Astra can make high-level decisions, such as knowing that plates go in a cupboard, while Eka's model handles the physical work of picking up a dish without breaking it. Haarnoja said most of the field approaches robotics from the top, starting with high-level planning, while Eka builds from the bottom up, starting with physical interaction.
Eka is running secretive pilots with industrial customers in pick-and-place and wire manipulation, but moving from lab to production line remains a major leap. Hauser said the big tests are running on real assembly lines without breaking for an entire year and showing the techniques work beyond pick-and-place. Eka's gripper costs less than $1,000 and is rated for 5 million duty cycles , short of the 20 million cycle standard set by industrial arms. Agrawal says the company is starting pragmatically with two verticals and plans to expand once they succeed.