Fern, a robotics startup backed by Y Combinator, said in early November 2026 that its new platform can convert three teleoperation episodes, in which a human remotely guides a robot through a task, into a high-fidelity simulation. The company says teams can then run thousands of evaluations without touching robot hardware. The figures come from the founders and have not been independently audited, according to RuntimeWire's coverage of the launch.
What Fern says it built
Fern operates under the corporate name Ishiki Labs, Inc. Its website describes the product as high-fidelity reinforcement-learning and evaluation environments for robots. Co-founder Robert Xu announced the launch in a post on X. The original could not be retrieved directly, though its text appeared on an indexed mirror and matched RuntimeWire's account.
The company's whitepaper names the problems it wants to solve: operator time, hardware wear, manual resets, inconsistent test conditions, and the infrastructure work required to scale simulations. Those are Fern's own descriptions. No independent study has put numbers on the costs involved.
It would also be a mistake to read the product as purely a neural world model, meaning an AI system that learns to predict how an environment behaves. Fern's blog documents a physics-simulation offering too.
What isn't known
The gaps are considerable. RuntimeWire said the report it could access did not identify a customer or give a measurement period. It also did not explain how "simulation" was defined, say whether anyone had used the product on paid or free terms, or cite any independent verification.
Fern has also described the requirement differently before. An earlier LinkedIn post from Xu, announcing a robotics-evaluation platform, cited a six-hour requirement. None of the sources reconciles that with the three-episode claim, and the two may simply reflect different approaches or product versions.
Fern's GitHub repository, fern-eval, contains integration examples involving Anvil Robotics and OpenArm hardware, plus a CloudChef example. The materials reviewed treat these as technical references. They are not signed customer contracts or endorsements.
Company and funding
Y Combinator's directory lists Fern as a Winter 2026 company, founded in 2025, with two team members, and names Gustaf Alstromer as its primary partner. Fern did not announce a new accelerator admission this month.
No Fern-specific financing details appear in the materials reviewed. Dealroom lists $125,000 from YC in January 2026, but that record carries an outdated product description and misnames the co-founder, so it is a shaky basis for any funding claim.
The company was building something else not long ago. Fondo published a post on Feb. 27, 2026, headlined "Fern Launches: Realtime Socially Aware AI for Meetings." A Fern robotics case study shows the company working in robotics by June 2026. When exactly the pivot happened, and what became of the meeting product, remains unverified.
Xu's co-founder is Amit Yadav, according to a YC LinkedIn post naming both. YC lists his education as Purdue University College of Engineering and IIT Gandhinagar. A Sept. 18, 2026, arXiv paper names him first among its authors on audio-model research. That is useful background about his work, though it says nothing about how Fern's robotics system is built.
A crowded field
Fern is entering territory with established players. NVIDIA offers Isaac Sim, its robotics simulation platform. Within YC, One Robot, founded by Hemanth Sarabu and Elton Shon, builds on robot-learning work at Industrial Next, and Runtime.si has a YC profile as well.
For now, interested teams can reach Fern at [email protected], the address published in its whitepaper. The next YC Demo Day, on Dec. 2, 2026, belongs to the Fall batch and not to Fern's.
