> ## Content Index
> Fetch the complete content index at: https://nexi.fund/llms.txt
> Use this file to discover other available public pages before exploring further.

# Eka's $1.45B Bet: Robot Hands That Feel Before They See
- URL: https://nexi.fund/eka-robotics-force-sensing-2026/
- Published: 2026-10-28T13:30:25.000Z
- Updated: 2026-10-28T13:30:26.000Z
- Description: Eka Robotics closed a $150M Series A at a $1.45B valuation on a contrarian thesis: robot manipulation improves faster by learning contact physics in simulation than by copying human video. The demo is striking. The duty-cycle math is not yet proven.
- Author: Nexi.fund Labs
- Tags: Defence & Robotics, #mode-4, #hook-question, #track-A, #layout-thesis-spine

What if the bottleneck in robot manipulation was never the camera?

In a lab in Cambridge, Massachusetts, a robotic hand picks up a raspberry without crushing it. The same hand screws in a light bulb, then pushes a wire into a connector. It is not seeing harder than other robots. It is feeling its way through the task.

On October 20, Eka Robotics closed a $150 million Series A at a $1.45 billion post-money valuation, led by Andreessen Horowitz and F-Prime. The round prices a contrarian claim: that contact is where general-purpose manipulation gets won, and that vision alone will not get there.

🎯

**The money is a bet on architecture.** Eka's vision-force-action model learns physics in simulation instead of copying human video.  
  
**Reported throughput is 1,500 picks an hour** at more than 95% success, roughly two and a half times the human baseline of about 600.  
  
**The unresolved number is durability.** The gripper is rated for 5 million duty cycles against the 20 million industrial benchmark, and its pilot customers are still unnamed. 

## Two schools of robot touch

The dominant pattern in robot learning is the vision-language-action model. A camera sees the scene, a language instruction sets the goal, and a policy maps both to motor commands. Physical Intelligence and Figure have built some of the most capable systems on that pattern. It works best when a task can be described in words and judged by sight.

Contact-rich work is the awkward case. A raspberry does not care how well a model recognises it. It cares how much force the fingers apply. Video of a human hand does not record that force, so a model trained on demonstrations has to infer it from pixels and hope the inference survives contact.

Eka's alternative puts force sensors in a custom gripper and feeds that signal into the policy alongside vision and language. When the grip begins to slip, the model changes pressure mid-task. The company calls the architecture vision-force-action, or VFA. The hand is designed in-house, with more force sensors than a standard gripper.

The two schools also disagree about where training data comes from. VLA systems lean on human demonstrations and video. Eka trains in a physics simulator with large-scale reinforcement learning, then transfers the policy to hardware. The founders argue that a simulator can generate millions of contact events, while a human video can label almost none.

> Eka's Vision-Force-Action model is the first approach I've seen that truly solves manipulation with the generality and throughput needed for deployments.— Raghu Raghuram, general partner, Andreessen Horowitz

That is an investor talking about his own deal. The useful test is what each approach can actually do on a line.

| Dimension           | VLA school                           | VFA school                            |
| ------------------- | ------------------------------------ | ------------------------------------- |
| **Training signal** | Human video and demonstrations       | Simulated contact and force feedback  |
| **Strength**        | Tasks that can be described and seen | Fragile, slippery and insertion tasks |
| **Weak point**      | Blind to force until contact fails   | Sim-to-real gap in contact dynamics   |
| **Evidence**        | Broad task demonstrations            | 1,500 picks/hour, above 95% success   |
| **Durability**      | Not the headline metric              | 5M cycles now, 20M target             |

Sources: Eka Robotics, Andreessen Horowitz, TechCrunch, October 2026

## The numbers behind the pitch

Eka was founded in 2024\. Pulkit Agrawal runs the Improbable AI Lab at MIT. Tuomas Haarnoja worked at Google DeepMind and co-created Soft Actor-Critic, a reinforcement-learning algorithm that many robotics teams still use. They met as graduate students at UC Berkeley.

$1.45B post-money valuation 

#### Series A price tag

$150M raised, led by a16z and F-Prime. *Eka Robotics, Oct 2026*

The seed round was $11 million in May 2025, led by Obvious Ventures and E14 Fund. With the Series A, investors have put about $161 million into the company. AI Magazine notes that Carta put the median 2026 Series A at $14.7 million. Eka's round is roughly ten times that median, which tells you the price is being set on a technical claim rather than current revenue.

The claim is specific. Eka says its robots pick objects outside their training distribution at 1,500 operations an hour. The human baseline it uses is about 600\. It handles objects up to 1 kilogram and succeeds more than 95% of the time. The same arm can seat a light bulb and insert a wire, two tasks with very different failure modes.

1,500 picks per hour 

#### Throughput claim

About 2.5x the roughly 600 picks a human manages. *Eka Robotics, Oct 2026*

The hardware number is the one industrial buyers will circle. The gripper costs less than $1,000 and carries more force sensors than a standard unit. That combination is unusual at the price. A force-sensing hand that costs five figures is a research instrument. One under $1,000 is a component you can put on many arms and replace when it wears.

WIRED saw an earlier version of the system in April and described its claw as close to a ChatGPT moment for the physical world. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, told TechCrunch the demonstration had a "wow factor" compared with other robot demos he had seen recently. He also noted that the arm moves fast enough to hurt itself, which implies the control stack is trusted not to destroy the hardware.

## What breaks in a full shift

A demo runs for minutes. A factory runs for years. Andreessen Horowitz's own write-up names the gap: contact dynamics are hard to model, and small errors in friction, compliance or actuator behaviour can turn a good simulated policy into a useless real-world one.

Eka's answer is that the model has already shown signs of generalisation across objects, tasks and environments. The public evidence for that is throughput and success rate in laboratory induction and bin-picking. The pilot customers remain unnamed, and the tasks disclosed so far are pick-and-place and wire handling.

Durability is the second gap. The gripper is rated for 5 million duty cycles. Standard industrial arms are built for 20 million. That is a fourfold difference, and it is the difference between a tool that survives one shift and a tool that survives a maintenance schedule. Agrawal has said the 20 million target is the goal, not the current spec.

#### What would count as proof?

A named customer running Eka hands through full production shifts, with uptime and cost per unit disclosed.  
  
Gripper life reaching the 20 million cycle benchmark in the field, not in a lab.  
  
Success rate holding above 95% on objects outside the training distribution after thousands of repetitions.  
  
A second hardware generation that keeps the sub-$1,000 cost while improving duty cycles. 

There is a useful comparison from the past week. [As we wrote in our piece on Enact's $30M pre-seed](https://nexi.fund/enact-intelligence-pre-seed-2026/), some robotics founders are attacking dexterity from the actuator upward. Eka is attacking it from the model and the hand together. Both are bets that the bottleneck sits in hardware-software co-design, not in a larger vision model.

## Why force is a data problem

Force sensing is not new. What is new is the price and the training loop. A research-grade tactile hand has historically cost far more than the arm it sits on, which kept force feedback in labs. Eka's sub-$1,000 gripper changes the unit economics of collecting force data: you can deploy many hands, gather contact events in parallel, and feed the misses back into training.

Simulation supplies the other half. Reinforcement learning can explore contact at a scale no human demonstrator can match, but only if the simulator's friction and compliance are close enough to reality. That is the sim-to-real problem, and it is where the VFA thesis either holds or collapses. A policy that learns in a clean simulator and meets a greasy metal part on a real line is the exact failure mode the approach has to survive.

The company's answer is co-design. Eka builds the hand, the sensors, the policy and the controller as one system, so the simulator is tuned against hardware it also manufactures. That is harder to scale than buying an off-the-shelf gripper. It also gives the model a cleaner signal than a policy that has to compensate for someone else's sensor noise.

📊

**Signals to watch over the next two quarters**  
  
A named pilot customer with a production-shift case study.  
  
Field data on gripper duty cycles against the 20 million benchmark.  
  
Whether the 95% success rate holds on objects outside the training set after wear.  
  
A second product line beyond pick-and-place and wire handling. 

## The factory buyer's math

Industrial automation has a brutally simple scoreboard: units per hour, error rate, uptime and cost per unit. A force-first hand wins only if it moves those four numbers on a real line, for a real customer, across a maintenance interval. The 1,500 picks an hour matters because it sets the labour comparison. The 95% success rate matters because a 5% failure rate on a fragile part can erase the throughput gain.

Integrators are the next gate. Most factory deployments are assembled by third parties who care about repeatability and spare parts more than model architecture. If Eka's hand needs a specialist to service, the sub-$1,000 bill of materials stops being the relevant number. The cost of ownership includes the integrator's time, the downtime and the replacement cycle.

That is why the unnamed pilots are the weakest link in the evidence chain. They prove that someone is willing to test. They do not prove that the economics survive a quarter of production. The company says the new capital goes to more tasks and faster customer deployments, which is the right use of the money if the durability curve cooperates.

## The valuation is a bet on the next 15 million cycles

At $1.45 billion, Eka is priced like a platform company. The competitive field is crowded and well funded. Physical Intelligence, Figure and Agility Robotics all have capital and deployment ambitions. Eka's differentiation is narrow and testable: it owns the gripper, the sensors and the policy, and it trains the policy on contact instead of imitation.

That vertical integration is also the risk. Every layer has to work at once. A brilliant policy on a gripper that wears out in months is a science project. A durable gripper with a policy that cannot generalise is a commodity. The round buys time to prove both layers together, before the next financing reprices the bet.

The number to watch is not the valuation. It is the duty cycle count on the hands that come back from those deployments. Eka has bought time to prove the hand. The factory will send the invoice in cycles, not dollars.

[ Eka Robotics raises $150M to scale its vision-force-action model The Robot Report's account of the round, the investor list and the 1,500 operations-per-hour claim. The Robot Report ](https://www.therobotreport.com/eka-robotics-raises-150m-scale-vision-force-action-model-robots/?ref=nexi.fund) 

The clearest trade-press summary of what Eka says its robots can do and who paid for the round.

[ Investing in Eka Robotics Andreessen Horowitz explains the VFA thesis, the sim-to-real problem and the economics it expects the model to change. Andreessen Horowitz ](https://a16z.com/announcement/investing-in-eka-robotics/?ref=nexi.fund) 

The investor's own statement of the thesis, including the contact-dynamics risk it has to underwrite.

[ Eka is worth $1.45B thanks to a physical AI model that has experts raving TechCrunch's reporting on the founders, the gripper cost, the duty-cycle target and the 95% success claim. TechCrunch ](https://techcrunch.com/2026/10/20/eka-is-worth-1-45-billion-thanks-to-a-physical-ai-model-that-has-experts-raving/?ref=nexi.fund) 

Contains the most concrete hardware numbers, including the 5 million versus 20 million cycle gap.