774: RFM-1 Gives Robots Human-like Reasoning and Conversation Abilities, with @JonKrohnLearns
Jon Krohn, host of the Super Data Science podcast, explains how Covariant’s RFM-1 could bring language-guided reasoning and conversation to industrial robot arms.
In this talk
- 1:37
Why robotics needs its own foundation models
Krohn frames RFM-1 alongside NVIDIA’s GR00T and Figure’s humanoid robot work. He argues that fixed industrial robot arms may make RFM-1 especially consequential because they are already common in factories, where small errors can create costly waste.
- 6:15
How RFM-1 works
Covariant trained RFM-1 using internet material and data from real-world robot deployments, including images, video, sensor readings, and robot actions. The 8-billion-parameter transformer can take these different data types as input and produce them as output, supporting tasks such as scene analysis, grasp planning, and predicting outcomes.
- 8:43
Language-guided robot tasks
RFM-1 can take natural-language instructions and ask people for guidance when a task is unclear. Krohn says this could let teams adapt robot behavior in minutes instead of the weeks or months often needed, though full robot-program orchestration still depends heavily on Python and C++.
- 10:00
Limits and potential applications
Learned world models give RFM-1 a basis for anticipating physical outcomes, but Krohn notes that it had not yet been deployed to Covariant customers. Its 512-by-512 image resolution and 5-frames-per-second rate also constrain how well it handles small objects and fast motion. He sees potential uses in manufacturing, logistics, agriculture, construction, waste management, and some health-care tasks.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.