Covariant RFM-1

2024-03-11 · Publisher : Covariant · Domain : Robotics

Covariant unveils RFM-1, a foundation model for warehouse robots trained on text, images, video, robot actions and sensor readings collected by its production fleet.

What changes

A single model combines language, vision, video and robot actions.
Training data comes from warehouse robots deployed in production.
The model stays proprietary, inside the Covariant Brain.
It foreshadows the 2024 wave of robot foundation models.

### Background On March 11, 2024, Covariant, a Californian startup founded by former OpenAI and Berkeley researchers (Pieter Abbeel, Peter Chen, Rocky Duan), announced RFM-1 (Robotics Foundation Model 1). The company describes it as a foundation model that gives robots a human-like ability to reason about their environment. ### Specifications - **Data**: text, images, video, robot actions and a range of numerical sensor readings, collected by warehouse robots running the Covariant Brain. - **Target uses**: order picking, kitting, depalletizing and item induction, Covariant's core logistics business. - **Availability**: closed model, embedded in the commercial offering; no weights or code released. Model size is not stated here for lack of a verifiable source. ### Why it matters RFM-1 is one of the first models to claim "foundation model" status for robotics based on production data rather than lab demonstrations. It shows the advantage of players that already operate a deployed fleet: every robot in service feeds the dataset. ### Key takeaway RFM-1 opens the 2024 race for robot "brains". A few months later, Amazon licensed Covariant's models and hired its founders.

Covariant RFM-1

What changes

Links