Covariant is building RFM-1 from warehouse robot data into robot foundation models
A March 11, 2024 RFM-1 launch ties Covariant Brain deployment data to an 8-billion-parameter robotics model built from warehouse manipulation trajectories.
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.
Covariant Brain deployments generate data on perception, grasp selection, object motion, placement and recovery from failed picking attempts. Covariant uses this warehouse data in RFM-1 training.
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.
Covariant says RFM-1 can interpret robot instructions, reason about scenes and support actions across manipulation tasks.
Have a robotics update Korthos should review? Send news, deployments, product releases, funding rounds, research, or media to [email protected] or reach out on X at @agkorthos.