About
Fern builds high-fidelity, action-conditioned robot world models and reinforcement-learning environments for policy developers, RL researchers, and robotics companies. Its differentiation is enabling customers to evaluate and improve robot policies in software, using production data and customer-specific physics, without repeatedly running physical hardware.
Market
Fern competes in physical-AI and robotics infrastructure, specifically learned simulation, robot-policy evaluation, and reinforcement-learning environments. Unlike conventional physics simulators such as Isaac Lab, MuJoCo, Genesis, and SAPIEN, Fern learns action-conditioned world models from real robot data and fits custom environments to a customer’s own embodiments and physics, enabling evaluation and RL training without consuming physical-robot time.
Fern targets robotics companies and robotics engineering teams developing production robot policies, especially organizations with proprietary robot embodiments and substantial teleoperation data. Its specific users include policy developers who need managed checkpoint evaluation and RL researchers who need learned-physics environments without tying up physical hardware.
At a Glance
Problem
Fern addresses a core bottleneck in physical AI: robotics teams still evaluate policies by running them on real robots, which consumes operator time, wears hardware, requires resets, and can take hours for a meaningful checkpoint test. Real-robot testing also makes comparisons unreliable because changing conditions and calibration drift prevent apples-to-apples measurement. Reinforcement learning is even harder on hardware because useful training requires thousands of parallel rollouts, while conventional simulators require extensive custom assets and hand-tuned physics for each new environment.
The killer use case is a robotics company that has a production policy and large volumes of teleoperation data but cannot iterate quickly without tying up its fleet. Fern’s whitepaper describes closing the loop on a customer’s rice-scooping policy in simulation, with predicted joint actions and generated frames closely matching the policy’s real-robot behavior.
Product / Service
Fern provides learned simulators and reinforcement-learning environments built from real robot data. Its action-conditioned world models take a starting image and a stream of robot actions, then generate future frames representing what would have happened on the physical robot. Because the model learns the physics from real data, teams can evaluate the same policy checkpoint they intend to deploy, without simulation-specific fine-tuning, hardware access, operator queues, or calibration drift.
The delivery model combines a managed evaluation platform for policy developers with custom world models for robotics companies. Customers can fit models to their own robot embodiments, tasks, environments, and existing teleoperation data; the resulting environments support benchmarking as well as Gym-style offline and online reinforcement learning. The intended benefit is faster, more reproducible policy development and a continual-learning loop in which production robots improve more frequently without consuming equivalent real-world hardware time.
Market
Fern operates in robotics infrastructure for physical AI, specifically learned world models, robot-policy evaluation, and reinforcement-learning environments. Its closest substitutes are conventional robotics simulators such as NVIDIA Isaac Sim, which provides physically based virtual environments for designing, testing, and training robots, and MuJoCo, an open-source physics engine used in robotics research and development. Fern’s differentiation is to learn a simulator from a customer’s real robot data rather than requiring teams to recreate every environment through hand-built assets and physics tuning. The evidence does not identify a definitive list of direct startup competitors.
The company appears to be early but has moved beyond a purely conceptual stage: Y Combinator lists Fern as an active Robotics Infrastructure company in the Winter 2026 batch, and Fern reports a customer production-policy case study. Its site also describes a managed platform, public benchmarks, and custom customer world models. No reliable evidence in the research discloses revenue, pricing, customer count, or whether the company is already profitable, so its commercial scale and revenue status remain undisclosed rather than safely classifiable as pre-revenue.
Founders & Leadership
Funding History
Y Combinator
Bessemer Venture Partners, Y Combinator
Recent News
Y Combinator’s 2026 robotics directory describes Fern as providing reinforcement-learning environments for robotics companies, powered by a foundational world model. It also says companies can plug in and test their robot policies in Fern’s environments.
Fern published a whitepaper on learned simulators for evaluating general robot policies. The company says it is building scalable evaluation and reinforcement-learning environments for robotics.
Oregon State coverage reports that Fern shifted toward robotics and collaborated with Jonathan Hurst, co-founder of Agility Robotics. The article also connects Fern’s work to teaching robots real-world tasks through Agility’s Digit humanoid robot.
Y Combinator’s company profile identifies Fern as a 2025-founded robotics company led by Robert Xu and Amit Yadav. It describes Fern’s product as reinforcement-learning environments for robotics companies powered by a foundational world model.
Active Roles
0No active roles right now.
Get notified when they postBusiness Model
Fern has a B2B model centered on a managed cloud evaluation platform and custom world models fitted to customers’ robot embodiments and teleoperation data. Public materials do not disclose specific pricing or whether revenue is subscription-, usage-, or services-based.