Eka Robotics

Eka Robotics is a robotics company based in Cambridge, Massachusetts, developing force-aware systems for dexterous manipulation.^[10][11] Its Vision-Force-Action (VFA) model combines images with feedback from robotic hands during physical contact.^[2] The company was co-founded by Pulkit Agrawal and Tuomas Haarnoja and spun out of Agrawal's laboratory at MIT CSAIL.^[3] SiliconANGLE reported that the founders started it in 2025.^[9]

Eka trains its manipulation policies with reinforcement learning in physics simulation instead of relying on human demonstrations, and it pairs the model with a custom gripper that carries force sensors.^[2][8] On October 20, 2026, it announced a $150 million Series A led by Andreessen Horowitz and F-Prime. TechCrunch reported that the round valued the company at $1.45 billion post-money.^[1][8]

Company and research background

Agrawal is an associate professor in MIT's Department of Electrical Engineering and Computer Science. His research group, the Improbable AI Lab, studies how robots learn perception, control, manipulation, and locomotion through experience.^[4][6] Haarnoja previously worked at Google DeepMind.^[3]

Haarnoja co-authored the 2018 Soft Actor-Critic paper with Aurick Zhou, Pieter Abbeel, and Sergey Levine. That work introduced an off-policy reinforcement-learning algorithm that optimizes expected reward together with policy entropy, and evaluated it on continuous-control benchmarks.^[5] It is part of the founder's research background, rather than a published specification of Eka's VFA implementation.

MIT described Agrawal's work on robotic hands and wrists as a move toward control based on forces rather than position alone. Its 2025 account also identified him as an Eka co-founder, before the company's October 2026 financing announcement.^[6]

The two founders first met as graduate students at the University of California, Berkeley.^[8][10] Haarnoja's Berkeley dissertation, Acquiring Diverse Robot Skills via Maximum Entropy Deep Reinforcement Learning (December 2018), was advised by Abbeel and Levine. It covered soft Q-learning, soft actor-critic and automatic temperature adjustment, tested on simulated and real robotic manipulation and locomotion tasks.^[14]

Both founders had earlier worked on transferring simulation-trained policies to real robots; MIT credits Agrawal's group with pioneering the use of simulation to train robot controllers that work in the real world.^[6] At Google DeepMind, Haarnoja was the first author of a 2023 paper on using deep reinforcement learning to train a small humanoid robot with 20 actuated joints to play a simplified one-versus-one soccer game. The agent was trained in simulation and transferred to real robots zero-shot.^[13] Agrawal and his co-authors Tao Chen and Jie Xu presented a model-free system that learned to reorient more than 2,000 geometrically different objects with a hand facing either up or down, demonstrated in simulation.^[10][12] The paper received the Best Paper Award at the Conference on Robot Learning (CoRL) in 2021.^[4] A 2023 follow-up in Science Robotics, "Visual Dexterity," described a real-time controller for reorienting novel and complex object shapes.^[4]

WIRED reported that Agrawal, who stayed in touch with members of OpenAI's Dactyl team, said that project's simulation approach had come to be seen as a dead end because of the so-called sim-to-real gap. He and Haarnoja, working in separate labs, remained convinced they could close that gap by making the simulation closer to reality.^[10] UC Berkeley professor Ken Goldberg, who has known Agrawal since his student days, is an adviser to the company, according to the same report.^[10]

Eka says its team brings experience from institutions and companies including MIT, Berkeley, DeepMind, Boston Dynamics, Microsoft, CMU, UPenn, Harvard, Northeastern, Caltech and Oxford.^[1] WIRED described its office and testing space in Kendall Square, Cambridge, a short walk from MIT.^[10]

Vision-force-action model

Eka designs its model, sensing, controller, and hardware together. The approach uses contact as information: slipping can prompt a grip adjustment, while resistance during insertion can prompt a change in motion. Vision provides information about the scene; force feedback provides information about interactions after contact begins.^[2]

The company describes generalization across objects, tasks, and environments as a goal for its VFA model.^[7] These are manipulation capabilities, rather than a claim that the system can carry out arbitrary physical work.

TechCrunch contrasted the design with vision-language-action models, which are trained to relate language, visual inputs and actions.^[8] "Vision is not the primary way in which you close this last millimeter problem," Agrawal told the outlet. "Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces."^[8] The outlet wrote that the promise of a model like Eka's is that it handles the physical side of a task while a more general model decides how to solve the problem. Haarnoja said most of the field has focused on planning and approaches robotics "from the top," pushing high-level intelligence down toward the physical, "whereas we think about it from the bottom up."^[8]

Training and transfer to hardware

Eka trains manipulation policies with reinforcement learning in physics simulation. The company says it varies objects, environments, and physical conditions to generate robot experience before transferring learned behavior to hardware.^[1] This makes sim-to-real transfer, particularly contact dynamics, a central part of its approach.^[2]

The company compares its method to AlphaZero, which reached superhuman play in chess and Go through reinforcement learning and self-play without human data. It says simulation lets it generate thousands of hours of robot experience in minutes, and argues that large-scale reinforcement learning buys performance while breadth of experience buys generality.^[1] Andreessen Horowitz argued that learning from human demonstrations (imitation learning) is limited because demonstrations take human time and rarely show a robot the many ways a task can fail.^[2]

WIRED reported that the model learns from a simulation that includes realistic joints and motors as well as physical principles such as mass and inertia. According to WIRED, the founders declined to give details of how they train their robots, since this is their commercial edge, and would not say exactly how they transfer learning from simulation to the real world.^[10] TechCrunch likewise said the founders were guarded about the details, beyond using a physics simulator to model the forces a robot meets while grasping and running that as a reinforcement-learning environment at scale.^[8]

Gripper hardware

Eka designed its own robotic hand. TechCrunch described it as more flexible than traditional grippers and fitted with more force sensors; it reacts to contact, which the outlet said is why it does not crush a raspberry.^[8] WIRED described custom grippers that incorporate a sense of touch, and photographed a newer machine with a three-point hand.^[10] For background on the hardware category, see dexterous hand and tactile sensing.

Agrawal told TechCrunch that the gripper hardware costs less than $1,000 and is rated for 5 million duty cycles. The company's goal is 20 million cycles, the bar set by standard industrial arms.^[8] A hardware test engineering role the company advertised in October 2026 covered validating actuators, grippers and full robotic arms, and leading certification work under ISO 10218 for collaborative environments and ingress protection.^[15]

Demonstrations and performance claims

The October 2026 announcements and earlier press coverage describe several manipulation tasks, with different levels of evidence:

Eka's own announcement says the robots pick objects outside the training distribution at 1,500 operations per hour, which it puts at roughly two and a half times a human baseline of about 600 operations per hour.^[1]

The throughput is a company-reported laboratory result, not an independently established production rate. Andreessen Horowitz identified maintaining performance throughout full production shifts as the next test. The picking figure should not be read as the rate for bulb installation or wire insertion.^[2]

Customers and deployments

Eka had not named any customers as of October 2026.^[9] TechCrunch reported that it was running pilots, which the outlet called secretive, with a number of industrial customers in pick-and-place and wire manipulation.^[8] Agrawal said the company was starting with two verticals, "pick, arrange, place, and wire handling," and planned to broaden its work once those succeeded.^[8]

The company's job board in October 2026 listed a forward-deployed robotics engineer role for setting up arms, grippers and camera systems at client sites, handling fleet issues, and writing deployment procedures for third-party integrators. All listed roles were in the Boston area.^[15]

Funding

On October 20, 2026, Eka announced a $150 million Series A led by Andreessen Horowitz and F-Prime. Obvious Ventures and E14 Fund, identified as its seed lead investors, continued their support.^[1]

The company also named MFV Partners, Bison Ventures, Glasswing Ventures, and Samsung Next as participants, alongside unnamed Fortune 500 strategic investors. Eka said the funding would support expansion of VFA to more tasks and accelerate customer deployments, with sustained reliability in real operations as its focus.^[1]

Eka's announcement did not state a valuation; the $1.45 billion post-money figure comes from TechCrunch.^[1][8] The company describes F-Prime as a venture capital fund affiliated with FMR LLC, the parent company of Fidelity Investments.^[1] Raghu Raghuram, the Andreessen Horowitz general partner who co-led the round, told TechCrunch that Eka was "already exceeding superhuman performance for certain tasks."^[8] SiliconANGLE put total investment in the company at about $161 million after the round.^[9]

Tech Funding News noted that the valuation was lower than the reported valuations of rivals it named, including Physical Intelligence, Generalist and Skild AI.^[11]

Reception

In a WIRED feature published in April 2026 after a visit to Eka, senior writer Will Knight wrote: "In more than a decade of writing about robots, I have never seen one move so naturally." He compared the robots to GPT-1, OpenAI's first large language model, which was often incoherent but showed glimmers of general linguistic intelligence. He also wrote that he did not know whether Eka's approach was the route to a ChatGPT-like breakthrough, noting that some experts believe mixing human demonstration with simulation will yield better results than simulation alone.^[10]

Kris Hauser, a professor at the University of Illinois Urbana-Champaign, told TechCrunch the robot had a "wow factor" compared with other recent demonstrations and "was able to adapt to a large number of different objects without retraining." He also named open questions: whether the system can run on real assembly lines with integrators "without breaking for an entire year," and how well the techniques generalize beyond pick-and-place.^[8] An unnamed roboticist at a frontier lab compared seeing the robot to the period when robots first achieved reliable legged locomotion.^[8] Tech Funding News wrote that the next proof point would be a named customer running Eka's robots on a production line for months.^[11]

Continue reading

• VVision-language-action modelA vision-language-action model (VLA model, or VLA) is a machine-learning model for robotics that uses visual observations and a natural-language instruction to generate actions a robot can execute.Listed in See also
• RRobot learningRobot learning studies how robots acquire or improve behavior from data and experience.Listed in See also
• LList of neolabsThis list of neolabs covers the 106 companies on "The Ultimate Neolab List", a list kept by Debarghya "Deedy" Das, a partner at the venture capital firm Menlo Ventures.Listed in See also
• NNeolabA neolab, also written neo-lab or NeoLab, is an informal term for an AI research company organized around a technical research program.Listed in See also
• RRobot foundation modelA robot foundation model is a large-scale machine learning model, typically based on the transformer architecture, that is pre-trained on broad, diverse datasets of robot interactions and then adapted to a…Listed in See also
• PPhysical AIPhysical AI is a contested umbrella term for artificial intelligence systems whose decisions are coupled to sensing and action in the physical world.Listed in See also

More to read

• Massachusetts Institute of Technology
• Robotics
• Pieter Abbeel
• Sergey Levine
• Humanoid Robot
• Imitation Learning

Topics:Physical AIRoboticsRobotics Companies

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Eka Robotics. "Eka Robotics Raises $150M to Build Superhuman Intelligence for the Physical World." October 20, 2026. Company announcement.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7Raghu Raghuram and Yoko Li. "Investing in Eka Robotics." Andreessen Horowitz, October 20, 2026. Investor announcement.
3. ^1 ^2 ^3Audrey Woods. "Startups at CSAIL: A Roundup of Recent Ventures." MIT CSAIL Alliances, August 31, 2026. University account.
4. ^1 ^2 ^3 ^4Pulkit Agrawal. "Pulkit Agrawal." MIT CSAIL faculty research page. Accessed October 21, 2026. Faculty profile.
5. ^1 ^2Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. "Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor." Proceedings of the 35th International Conference on Machine Learning, PMLR 80:1861-1870, 2018. Research paper.
6. ^1 ^2 ^3 ^4Alex Shipps. "Pulkit Agrawal receives the IROS Toshio Fukuda Young Professional Award." MIT CSAIL, October 24, 2025. University research background.
7. ^Eka Robotics. "Eka Robotics." Official website. Accessed October 21, 2026. Company overview.
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17Tim Fernholz. "Eka is worth $1.45B thanks to a physical AI model that has experts raving." TechCrunch, October 20, 2026. News report.
9. ^1 ^2 ^3 ^4 ^5 ^6Duncan Riley. "Eka Robotics nets $150M at a $1.45B valuation for robot hands that sense force." SiliconANGLE, October 20, 2026. News report.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Will Knight. "I've Covered Robots for Years. This One Is Different." WIRED, April 29, 2026. Feature article.
11. ^1 ^2 ^3Abhinaya Prabhu. "Eka Robotics raises $150M at $1.45B valuation led by a16z and F-Prime for dexterous robots." Tech Funding News, October 21, 2026. News report.
12. ^Tao Chen, Jie Xu, and Pulkit Agrawal. "A System for General In-Hand Object Re-Orientation." Conference on Robot Learning (CoRL), 2021. arXiv:2111.03043. Research paper.
13. ^Tuomas Haarnoja, Ben Moran, Guy Lever, et al. "Learning Agile Soccer Skills for a Bipedal Robot with Deep Reinforcement Learning." arXiv:2304.13653, April 26, 2023. Research paper.
14. ^1 ^2Tuomas Haarnoja. "Acquiring Diverse Robot Skills via Maximum Entropy Deep Reinforcement Learning." PhD dissertation, Technical Report UCB/EECS-2018-176, EECS Department, University of California, Berkeley, December 14, 2018. Dissertation record.
15. ^1 ^2Eka Robotics. "Jobs at Eka Robotics" (including "Forward Deployed Robotics Engineer" and "Senior/Staff/Principal Hardware Test Engineer" postings, October 2026). Ashby job board. Accessed October 27, 2026. Job board.

Show all 15 references