Covariant is building RFM-1 from warehouse robot data into robot foundation models

Covariant introduced RFM-1 on March 11, 2024, positioning the system as a robot foundation model trained on warehouse manipulation data. TechCrunch framed the work as a ChatGPT-style model for robots, while IEEE Spectrum reported that the model used tens of millions of trajectories from production warehouse systems and roughly 8 billion parameters.

The warehouse data is the center of the story. Covariant Brain has been deployed in robotic picking and item-handling cells where the work creates repeated examples of perception, grasp selection, object motion, placement, and recovery from failed attempts. That gives RFM-1 a stronger operating base than a model trained only on curated lab tasks or internet video.

Covariant was founded in 2017 by a team that included Peter Chen and Pieter Abbeel, with roots in UC Berkeley robotics research. The company first built commercial traction around robotic picking for warehouses, where variation in SKUs, packaging, bins, and presentation creates a useful training surface for general manipulation.

RFM-1 broadens that thesis from picking software into physical AI infrastructure. Covariant says the model can understand robot instructions, reason about scenes, and support actions across manipulation tasks. The strategic shift is from optimizing one warehouse cell toward packaging learned robot behavior as a reusable model layer.

The competitive field includes Skild AI, Physical Intelligence, Sereact, Google DeepMind robotics work, NVIDIA GR00T and Isaac tooling, Amazon's internal robot-learning programs, and warehouse manipulation companies building proprietary task models. Covariant's distinction is production manipulation data from deployed systems rather than only foundation-model ambition.

The proof boundary is transfer. Public material does not disclose broad performance outside warehouse manipulation, customer retention by application, task-level benchmark results, fine-tuning cost, or how much integration is needed when RFM-1 moves into a new robot body or workflow.

If RFM-1 turns warehouse manipulation experience into reusable robot behavior, Covariant can move beyond picking automation into the intelligence layer for machines that need to handle physical variation. The company is betting that real robot work generates the data advantage foundation models need.

Have a robotics update Korthos should review? Send news, deployments, product releases, funding rounds, research, or media to tips@korthos.xyz or reach out on X at @agkorthos.