Three things I read this week on Physical AI, with a short note on each and a link to the source. The notes are mine. The reporting is theirs, so go and read it.
A robot model that predicts video and actions together, now open
Runway released Praxis-1, which it calls a world action model, with open weights under the Apache 2.0 license. It is a 27-billion-parameter model that predicts video and robot actions together. The company says partners have run it on three very different machines: a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. The Robot Report
My note: The design choice I noticed is that video prediction stays in the training objective the whole way through. The model has to picture what happens next before it moves. Brains work in a similar way: a large part of perception is prediction. This is also the point where generative models and physical ones meet, which I unpack in Physical AI vs Generative AI.
Teaching robots with hundreds of years of video
In an interview, Rhoda AI’s chief executive Jagdeep Singh described how the company’s direct video action model works. He says it is pretrained on hundreds of years of broad internet video, then given roughly 10 to 20 hours of training on the specific task. The model predicts about half a second of video ahead, a second stage turns that prediction into robot actions, and the loop repeats with a fresh observation. The Robot Report
My note: Two timescales in one system: a slow, broad prior and a short adjustment for the job at hand. The half-second loop is the part I would study. Predict, act, look again, so the loop closes on the world and not on the model’s own guess. If the vocabulary is new, start with vision-language-action model in the glossary.
A warehouse cart you can push around
MIT News profiled Robust AI, co-founded by Anthony Jules and MIT Professor Emeritus Rodney Brooks. Its mobile robot, Carter, looks like a utility cart and has a force-sensitive handlebar, a vision system and wheels that move in any direction, so a person can push, pull or redirect it with one hand. The robots are deployed in warehouses and on factory floors, where they carry items across large facilities. MIT News
My note: A handlebar that senses force is a humble sensor doing an important job. It lets a person and a machine negotiate through touch, the way two people carry a table without discussing it. Nobody reads a manual to push a cart.
More: all Physical AI news notes · what Physical AI is.