Eka Robotics bets on force, not language, to teach robots dexterity

Eka Robotics has emerged from stealth with a Vision-Force-Action model that it says can push robots beyond the long-standing trade-off between generality and speed in manipulation tasks. The Cambridge, Massachusetts startup, was co-founded in 2025 by MIT’s Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja. The deep tech entrepreneurs are pitching force sensing and simulation as the route to more capable machines.

In robotics, much of the recent excitement has centred on vision-language-action systems, which treat language as a bridge to physical control. Eka says that is too indirect for the contact-rich realities of the physical world. Its approach instead tries to make robots learn mass, friction and inertia through practice in high-fidelity simulation, then transfer those skills to the messier settings of factories and homes.

Across the robotics industry, the race is on to build foundation models that can scale across tasks, rather than brittle systems tuned for one environment, and the prize is a larger share of warehouse work, light manufacturing and household assistance. The strategic question is whether the winning path is imitation from human data, reinforcement learning in the real world, or simulation-first training that seeks to compress years of trial and error into computational time.

“We’re building intelligence for the physical world in its native language: forces,” Pulkit Agrawal wrote on LinkedIn. In the same post, he added that robotics has long faced a trade-off between “generality” and “speed,” and that “the real world requires both”.

Eka’s presentation suggests confidence that force-aware control can do more than sort objects or pick up toys. The company has highlighted tasks such as screwing in a light bulb and handling slippery items, small feats that still define the frontier of robotic manipulation. For now, the message is as important as the model: the next leap in robotics, Eka is arguing, will come not from making machines more verbal, but from making them more physical.