Models trained around warehouse work
Covariant developed AI systems for robotic picking and other warehouse manipulation. Its approach used real robot interactions alongside internet data, with the aim of transferring experience between objects, tasks and robot configurations. The company's RFM-1 account describes inputs and outputs spanning language, images, physical measurements and robot actions. [1]
The March 2024 RFM-1 introduction offered live demonstrations at a trade show and by appointment. The developer described predicting consequences of actions and accepting language instructions. This established a research and commercial demonstration, rather than a public download of weights or an unrestricted robot-control API. [1]
A license and team move to Amazon
Amazon subsequently described a non-exclusive license to Covariant's models and the hiring of three founders and part of the team. Its announcement explicitly said Covariant would continue serving its own customers. A technology license and staff move do not establish that Amazon bought the company, so that agreement is not represented as a parent relationship. [2]
Sources
- Covariant Introduces RFM-1 to Give Robots the Human-like Ability to Reason
Covariant · Mar 11, 2024
- An update on how we’re accelerating the use of AI in robotics at scale
Amazon · retrieved Oct 23, 2026