Despite heavy spending and vast amounts of video data used to train deep learning systems, robots still face challenges when grasping objects.

Eka Robotics aims to shift this situation with a manipulation system that pairs a force-sensitive gripper and a novel deep-learning approach. The company, started by MIT professor Pulkit Agrawal and ex-Google DeepMind researcher Tuomas Haarnoja, recently secured a $150 million Series A round. This funding values the firm at $1.45 billion post-money, with a16z and F-Prime leading the round.

Raghu Raguram, an a16z general partner who helped lead the round, noted that the system already surpasses human performance on select tasks. He expects superhuman results across many dexterous manipulation jobs within months.

In a field full of robotic showcases, Eka's performance stands out to experienced builders. University of Illinois Urbana-Champaign professor Kris Hauser highlighted its strong impact compared to recent demos. The robot adapts to many objects without retraining and shows a human-like persistence when initial attempts fail, he added.

One roboticist from a frontier lab, speaking anonymously, described the demo as moving, likening it to early successes in legged locomotion years ago. The motion appears notably fluid and quick. The arm manages roughly 1,500 objects up to 1 kg each per hour, outpacing a human's typical 600, while achieving over 95 percent reliability. Its speed also means it could damage itself, signaling deep confidence in the control algorithm.

During a test, the roboticist offered the arm a raspberry. The sudden approach followed by a gentle slowdown and light contact left a lasting impression, they recalled.

How the system avoids damaging delicate items

The results stem from the model's design. Instead of vision-language-action frameworks, Eka uses a vision-force-action model centered on force data and actions. This aligns with views from critics like Rodney Brooks, who argue that force and touch representation are essential. Agrawal emphasized that vision provides only rough positioning, while force sensing handles the final grasp.

The team built a custom hand with added flexibility and sensors that respond to contact. Training relies on physics simulation to model forces during grasping, creating scalable reinforcement learning setups. This generates thousands of hours of experience rapidly, allowing the robot to exceed human demonstration limits, as Hauser observed.

Agrawal and Haarnoja, who met at UC Berkeley, focus on scaling reinforcement learning through simulation rather than lengthy human-collected videos. This choice supports reliable object interaction, a longstanding hurdle.

Agrawal stressed that the approach closes the gap between broad capability and high performance seen in many current systems.

Turning versatile robots into practical tools

Highly general physical AI models often lack precision on specific jobs, while narrow models may not support strong business cases. Eka's model targets physical execution, leaving high-level planning to broader systems. For instance, it can manage fragile items like dishes without breakage.

Haarnoja explained the bottom-up focus on physical control over top-down planning. The company runs private pilots in pick-and-place and wire tasks with industrial partners, yet real-world deployment remains a key test.

Hauser noted challenges in sustaining operation on assembly lines for extended periods and expanding beyond pick-and-place. Agrawal aims for hardware durability, targeting 20 million cycles with the current gripper already rated for 5 million at under $1,000 cost.

The team plans to refine workflows in select verticals first, then expand the method further, Agrawal said.