Skip to main content
← engineering

One Robot Builds World Model Eval Platform for Manipulation Policies

By Zero G Talent•

Training end-to-end manipulation policies for robotics currently relies on a repetitive loop: engineering teams collect data, train a policy, deploy it onto physical hardware, identify points of failure, and collect more data. One Robot aims to replace this trial-and-error process with task-specific world models and an evaluation layer designed to pinpoint failure modes before hardware deployment.

Founded in 2025 and based in San Francisco, the company raised a $500K Seed round with backing from Y Combinator (Batch W26), Accel, Toloka.vc, and angel investors across the robotics and artificial intelligence sectors. The team operates as a deliberately small group where all members work as individual contributors handling wide technical surface areas across model controllability, long-horizon autoregressive generation, and sim-to-real validation.

Technical Background and Leadership

One Robot was co-founded by Hemanth Sarabu and Elton Shon, who previously led robot learning together at Industrial Next.

Sarabu, who serves as CEO, previously engineered perception and robot learning systems at Symbio Robotics, Google, and NASA JPL, and earlier bootstrapped geospatial AI firm Crescer AI to profitability. Shon, serving as CTO, brings a background spanning software, firmware, and distributed machine learning systems, having contributed to Tesla's Dojo supercomputing platform and Industrial Next's control infrastructure.

Together, the founders are addressing long-horizon generative tasks and controllability in simulated environments to ensure synthetic validation reliably transfers to real customer robotic arms.

Core Architecture and Hiring Priorities

One Robot is scaling its San Francisco engineering group across four technical openings, each tied directly to different components of its training and verification stack.

The team is hiring a Founding Machine Learning - World Models engineer ($150,000–$275,000) to build generative world models featuring action conditioning and high physical fidelity to simulate manipulation scenes. The role involves managing multi-GPU clusters, owning training infrastructure, and establishing a continuous data engine that compounds across customer-specific tasks.

To close the loop between prediction and physical deployments, the startup is also recruiting for Founding Machine Learning - Eval Layer ($150,000–$275,000). This engineer will train evaluation models and vision-language models (VLMs) that classify and verify policy behaviors, giving customers automated failure detection prior to factory production, supported by a self-improving engine informed by real-world interventions.

Connecting these foundation layers to physical hardware, the Founding Robot Learning role ($150,000–$275,000) focuses on training vision-language-action (VLA) models and end-to-end imitation policies for deployment on physical manipulators. This engineer tests world model fidelity against customer data to address sim-to-real transfer under production pressure. For early-career researchers, the onsite Machine Learning Intern role ($6,000–$8,000 per month) contributes directly to research in world modeling and policy evaluation, including hardware data collection and physical robot benchmark tests.

Open Roles at One Robot

All positions are based on-site in San Francisco, California.

Role Location Work Mode Salary
Founding Machine Learning - World Models San Francisco, CA, US Onsite $150,000 – $275,000
Founding Machine Learning - Eval Layer San Francisco, CA, US Onsite $150,000 – $275,000
Founding Robot Learning San Francisco, CA, US Onsite $150,000 – $275,000
Machine Learning Intern San Francisco, CA, US Onsite $6,000 – $8,000 / month

See all open roles at One Robot