Robotics finally gets a shared starting point instead of starting from scratch
A downstream-ready foundation model trained on 100,000 hours of manipulation data changes the default for new robotics projects. Language modelling has had a shared starting point for years; embodied AI has not, and every team has paid the data cost separately.
The asymmetry has been stark. A team building a language product starts from weights someone else spent a fortune training. A team building a manipulation system has started from nothing, collecting trajectories on its own hardware in real time — a cost that has kept robotics a capital-intensive field with few entrants.
Real-world trajectories are the expensive ingredient and they cannot be shortcut. Simulation helps but transfers imperfectly, and a hundred thousand hours of physical manipulation is a data asset that most organisations could never assemble. Releasing a model trained on it is the transfer of that asset to the field.
What to watch is generalisation across bodies. The field's recurring disappointment is policies that work beautifully on the platform they were trained on and fail on anything else. Embodiment-free pre-training is aimed directly at that problem, and independent replication on unfamiliar hardware is the test that matters.
Radical Data Science — AI news briefs bulletin board for August 2026 → · NVIDIA — National Robotics Week — physical AI research →