Eka Robotics, founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, has closed a $150 million Series A round that values the company at $1.45 billion post‑money. The financing was co‑led by venture firms a16z and F‑Prime.
A different approach to robotic manipulation
Rather than relying solely on vision‑centric models, Eka combines a force‑sensitive gripper with a vision‑force‑action deep‑learning architecture. The system ingests visual inputs and direct force feedback, and uses that combination to decide and execute grasping and manipulation actions. The hardware is a more flexible gripper than typical industrial end‑effectors and includes additional force sensors so the robot can react to contact and avoid crushing delicate objects like raspberries.
Why force matters
Pulkit Agrawal has argued that vision alone only gives approximate information about an object’s location and pose; closing the “last millimeter” of grasping requires force feedback. Critics of vision‑only robotics models — such as Rodney Brooks — have warned that touch and force must be represented for reliable manipulation. Eka’s results provide evidence that incorporating force inputs can materially improve dexterity.
Simulation and reinforcement learning at scale
Eka trains its models in a physics simulator that models the multiple contact forces a robot will encounter. Using reinforcement learning in that simulated environment, the company says it can generate thousands of hours of robot experience in minutes. This allows the robot to learn strategies that can exceed the quality of human demonstrations used for training.
Reported performance
Eka reports that its robot can pick about 1,500 objects per hour (objects up to 1 kg), compared with roughly 600 items per hour for a human, and that it reliably grasps objects more than 95% of the time. The arm’s speed and force are high enough that physical damage to the robot is a risk, which observers say indicates significant trust in the control algorithms developed during testing.
Industry pilots and deployment challenges
The company is running secretive pilot projects with industrial customers on pick‑and‑place and wire‑handling tasks. Moving from lab demos to continuous operation on production lines is a major challenge: experts note that the system will need to run for long periods (for example, years) without failures to meet industrial standards. Another open question is how well the techniques generalize beyond pick‑and‑place to more complex, value‑adding assembly tasks.
Hardware economics and lifecycle targets
Agrawal has said Eka’s gripper hardware costs under $1,000 and is currently rated for 5 million duty cycles; their longer‑term target is to reach the 20 million cycle benchmark typical of standard industrial robotic arms. The company is taking a pragmatic, vertical‑first commercialization strategy: focus on hitting home runs in two verticals (pick, arrange, place, and wire handling), then expand the approach to broader applications.
What this could mean for robotic AI
Eka’s combination of force‑aware hardware, simulation‑driven reinforcement learning, and a vision‑force‑action model addresses a key gap in many modern robotics systems: the generality‑performance tradeoff. If the approach proves robust in industrial deployments, it could enable more generally capable physical AI systems that handle the low‑level physicality while higher‑level models provide task planning and decision making. However, the leap from promising demos to durable, widely deployed systems remains to be demonstrated.
