One Robot

World models for robot evals and training.

Product Overview

One Robot sells simulation environments built on world models learned from real robot experience rather than from hand-authored scenes.

Its workflow runs in three steps: training a world model on customer robot data to learn physics and visuals, rolling out the customer policy in simulation to surface edge cases before deployment, and generating targeted synthetic trajectories to correct the failures found. The environments cover contact-rich manipulation, deformable objects, and precise assembly, and the company contact address is [email protected].

Market Outlook

The company frames robot-policy iteration as bottlenecked by real-world data collection and by access to robot time.

It positions simulation environments as the way teams run more training and evaluations and find failure modes faster. It names robotics teams that train and evaluate manipulator policies as its buyer set.

Competitive Advantages

The offering is built around learning the simulator from customer robot data rather than hand-authoring scenes.

That approach targets the two gaps the company names in existing simulators: visual realism and interaction realism. It is aimed at teams building action models for manipulators whose iteration is limited by how much physical robot time they can get.

Competitive Disadvantages

The company is early stage, with a two-person founding team and a single-page public website.

Its own description states that it is starting with hard manipulation tasks such as textiles and box folding, a narrower initial scope than general-purpose physics simulators that also cover navigation and locomotion.