Co-authored by Dionis Minev, Fine Structure Ventures

Robotics has a long history of automating mostly manufacturing and logistics tasks through delivery of highly repeatable and reliable operations that meet or exceed human performance. Much of this progress is from integrator-led deployments in highly deterministic environments where engineers meticulously hand-code robot movements. However, this approach has failed for innumerable tasks where the movements are too complex, and the environment too uncertain for traditional robotics to be relevant.

Today’s era of general purpose robotics offers the promise of automation to a vastly expanded set of activities by enabling robots to operate in unstructured environments without a pre-determined set of movements. Tasks like routing and connecting wiring harnesses in a car, seating thin and flexible cables inside electronic devices, or packing produce, where every item has a different shape and blemishes easily. While progress toward extending robotic capability to these people-centric tasks is remarkable, the key tenets of value creation—speed and reliability—have remained elusive. Achieving this would open general purpose robotics to the more than $2T spent on manufacturing labor alone in developed markets.

We think the missing ingredient to delivering economic value is giving a robot the ability not only to ‘see’ its environment, but also to ‘feel’ it through force. Eka Robotics solves general purpose robotics through its vision-force-action (VFA) model to deliver unparalleled speed and reliability, on top of the adaptability so many other start-ups have already proven. We are excited to partner with a16z and the team at Eka Robotics to finally unlock superhuman robotic capabilities for the next generation of general-purpose robots.

What happens when you can see but can’t feel?

Evolution has equipped humans with the ability to perform remarkably dexterous tasks without constantly looking at their hands. That ability depends on a continuous stream of force feedback from our hands and limbs that lets us feel what we’re doing and adapt on the fly.

Touch is easy to take for granted, yet humans depend on it for almost everything they do. As Rodney Brooks has pointed out, experiments show that when people’s fingertips are anesthetized, even a simple task like lighting a match becomes slow and uncoordinated. While vision guides the high-level plan, touch enables the high-frequency contact feedback required for dexterous manipulation. The importance of the hands also shows up in cortical maps of the brain that process sensory and motor functions, where they occupy a large share of the map.

Rare sensory neuropathies act as a permanent and more extensive version of the above mentioned anesthesia experiment. They can affect the whole body rather than just the fingertips, leaving a person unable to feel touch or sense where their limbs are. These patients have to relearn how to move using vision alone, watching their limbs and consciously planning every movement. While there are a few successful cases, most struggle to regain full mobility. Notably, in the successful cases, if attention drifts or the line of sight is interrupted, movement and manipulation fall apart. What these patients have lost is the fast, real-time reaction that biology gives us, and that we never have to explicitly think about.

Today’s robotic foundation models excel mainly at planning

Most general-purpose robot models remind us of the above, they rely predominantly on vision, planning their movements ahead and reacting slowly once they’re underway. The dominant approach is the vision-language-action model (VLA), which takes in camera images along with audio or text instructions and outputs a short sequence of actions, known as an action chunk. That chunk is then executed largely open-loop, with little ability to react or adjust while the movement is underway.

This planning layer has seen incredible progress over the last 24 months, but it is also getting increasingly commoditized. Several capable VLAs now have open-sourced their models, including Physical Intelligence’s π0.5, NVIDIA’s GR00T N1.7 and MolmoAct2 from the Allen Institute for AI. Proprietary models from companies like Skild, Generalist, and Dyna are still considered state of the art, but the gap between these models and open-source is shrinking. At the same time, general-purpose models like OpenAI’s Astra and Google’s Argon are starting to show they can operate robots directly, competing with VLAs. With so many players, a lasting edge is getting harder to build.

Eka recognized that planning is only half the problem, while the other half is controlling movement in real time, at the speed and reliability commercial use requires, which planning-focused models still struggle with.

The overlooked controller

Every planned action chunk is a guess about the world. But that guess can turn out wrong when an object turns out heavier than expected, or the friction between two surfaces differs from what the model assumed. A robot needs to sense these changes in contact forces and correct for them at a high frequency. That ability to react when the initial guess is wrong is what recovers a movement that would otherwise fail. It is also what current VLAs lack, and it is the source of the reliability and speed that separates current VLA demos from a commercially deployable product.

The challenge for force-aware control is striking the right balance between cost, complexity, and measurement fidelity. For years, roboticists have attempted to integrate different types of tactile sensors into their deployments. There have also been many innovations in end-effectors for dealing with more complex manipulation tasks, with a recent explosion of hand-like grippers. While technology in these domains is rapidly advancing, cost is a constant hurdle, high degree of freedom grippers are prone to failures, and companies building robot foundation models are hesitant to train their models on proprietary solutions.

Eka has solved this tradeoff through their VFA model, which combines reinforcement learning (RL) training in simulation for inferring forces, using a proprietary low-cost and reliable approach for force sensing. Early demos have been built with simple two-fingered grippers with articulated fingers, but the approach extends to multiple gripper form factors without relying on complex hand geometries.

Learning by search and exploration

Many of the most impressive results in AI have come from RL. Instead of copying examples, the policy tries things, sees what works and keeps searching. That kind of search can find strategies no human showed it. We have seen this successfully deployed in the digital world, from AlphaGo’s unexpected moves in the game of Go to today’s reasoning models applied to coding and math. In simulation, a robot policy can go through the same process of search and exploration, ending up in situations no one planned for and learning how to recover from them.

Eka has applied this approach to learning forces and mastering contact-rich manipulation. Unlike most of the field, Eka doesn’t rely primarily on human demonstrations, whether collected through teleoperation, handheld data-collection grippers, or video. Demonstrations are expensive to collect, scale linearly with the number of people collecting the data, and mostly show the task going right. Training in simulation scales with compute rather than human hours, and the resulting policies are fast, robust, and able to recover from mistakes. Of course closing the simulation to reality gap (sim2real) to make this work is difficult, but it is also where much of Eka’s know-how sits and what made the company’s dexterous demonstrations possible.

Eka Robotics

Eka launched publicly in April 2026 with its Vision-Force-Action model. Force is not just an extra input added to a vision-language model—the system is architected to natively understand forces. Eka uses simulation as the primary source of data, overcoming the sim2real gap, and the skills the policy learns transfer directly to real robots.

Pulkit Agrawal and Tuomas Haarnoja are the two of the most talented roboticists in the world and their research has laid the foundation for building Eka. Pulkit is a professor at MIT, where he runs the Improbable AI Lab and has spent years teaching robots dexterous skills. Tuomas comes from Google DeepMind and wrote Soft Actor-Critic, one of the most widely used RL algorithms. Few teams combine their depth of skill in reinforcement learning, robot simulation, technical know-how, and outside the box thinking. We couldn’t be more excited to be along for the journey.

Originally published on LinkedIn