A hundred thousand hours, given away

Language modelling has had a shared starting point for years. Robotics has had every team paying the data cost separately, in real time, on their own hardware. That asymmetry just broke.

A robot foundation model trained on more than 100,000 hours of real-world manipulation trajectories, released downstream-ready. The number is the story. Manipulation data cannot be scraped and cannot be parallelised — it accrues at one second per second, on physical hardware.

Why this changes the entry cost

A language team starts from weights someone else paid for. A robotics team has started from nothing, which is why the field has few entrants and heavy capital requirements. Turning every project from a data-collection campaign into a fine-tuning exercise is precisely the shift that moved NLP from labs into products.

The claim that needs testing

Embodiment-free pre-training plus real-robot grounding targets the field's oldest disappointment: policies that work beautifully on their training platform and fail on anything else. Whether transfer holds on hardware Xiaomi did not choose is the question, and independent replication is the only thing that answers it.

And a safety property arriving alongside

Systems that translate language into executable scripts rather than directly into motor commands create an inspection point that end-to-end policies lack. For deployment near people, a reviewable artefact between intent and motion is worth more than architectural elegance.

Robotics spent a decade short of data and short of trust. This week it got a down payment on both.

Radical Data Science — AI news briefs bulletin board for August 2026 → · NVIDIA — National Robotics Week — latest physical AI research →