Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Eka Robotics raises $150M at a $1.45B valuation for a force-first manipulation model

Eka Robotics raised a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime. The MIT spinout pairs a force-sensitive gripper with a vision-force-action model that it says picks about 1,500 objects an hour with better than 95% reliability.

Published
Eka Robotics 完成 1.5 亿美元 A 轮,14.5 亿美元估值押注力觉操作模型
Image source: techcrunch.com

Eka Robotics has closed a $150 million Series A at a $1.45 billion post-money valuation, TechCrunch reported on October 20, with a16z and F-Prime leading the round.

The startup was founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, who met as graduate students at UC Berkeley. Its bet is that robot manipulation has been held back less by planning than by touch.

Instead of the vision-language-action models that dominate robot learning, Eka built a vision-force-action model paired with a force-sensitive gripper that reacts to contact. Vision is not the primary way to close the last millimeter, Agrawal told TechCrunch: vision locates an object and reads its pose, while grasping and searching happen through force.

The numbers Eka quotes are striking. The robot picks about 1,500 objects weighing up to 1 kg per hour, against roughly 600 for a human, and the company claims better than 95% reliability. The arm moves fast enough that it could damage itself, which one frontier-lab roboticist read as confidence that the control algorithm will not break the hardware.

Training relies on simulation rather than videos of human demonstrations. Eka models the multiple forces a grasp encounters inside a physics simulator, turns that into a reinforcement-learning environment that runs at scale, and says it can generate thousands of hours of robot experience in minutes.

Outside roboticists are unusually warm. Kris Hauser of the University of Illinois Urbana-Champaign said the robot adapted to a large number of objects without retraining and showed a humanness in retrying after failure, while a16z general partner Raghu Raguram said it already exceeds human performance on some tasks.

Commercial proof is still missing. Eka is running undisclosed pilots with industrial customers on pick-and-place and wire handling, and Hauser notes that the hard part is surviving a full year on a real assembly line, then generalizing beyond picking to tasks that add more value.

On hardware, Agrawal says the gripper costs under $1,000 and is rated for 5 million duty cycles, short of the 20 million typical of industrial arms. The plan is to make two verticals work first, then extend the approach from there.

Why it matters

If the force-first approach holds up on real lines, robotics competition shifts from generalist planning toward reliable hands, a bet the $1.45 billion valuation already prices in.

Eka RoboticsRoboticsFunding
Back to realtime news

Nearby Updates

All

10/20, 20:56

DataGroomr launches an AI agent for natural-language Salesforce data quality

DataGroomr introduced an AI agent that lets Salesforce teams manage data quality through natural language, according to a PR Newswire release. The announcement did not disclose pricing, availability or the models behind the agent, leaving its practical scope unclear.

10/20, 20:02

Altrio launches AI agent that screens thousands of listings against investment criteria

Altrio has launched an AI agent that scans thousands of listings to find the ones that fit a customer's investment criteria. The product targets the first-pass screening stage of investment research, but its market coverage, data sources and accuracy have not been disclosed.

10/20, 19:33

Report: Tencent leases 100,000 AI chips from Oracle to work around US export rules

Startup Fortune reported that Tencent is leasing 100,000 AI chips from Oracle as a way around US restrictions on chip sales to China. The arrangement replaces direct hardware purchases with rented compute, though chip model, deployment location, and contract terms are not addressed.

10/20, 19:03

DeepSeek open-sources DeepInfer, the engine behind its production serving

DeepSeek has published DeepInfer on GitHub, the high-performance inference framework it uses to serve its models at scale, with the release focused on offline inference. The code covers optimized kernels, parallelism strategies, deterministic sampling and structured outputs, giving outside teams their first direct look at how DeepSeek actually runs its models.