A preprint posted to arXiv on 1 September at 17:58 UTC describes Facet-0, which its authors call a robotic foundation model that “predicts and values the contact consequences of its actions”. On five sub-millimetre computer-assembly tasks the task-adapted system reaches 82% mean success, against 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.
What the model actually predicts
Not pixels. The architecture is built around a joint action-wrench proposal — a wrench being the combined force and torque at the wrist. A causal wrench history is aligned with vision-language semantics and kinematic state, and flow matching generates each action chunk together with the future wrist-wrench profile it is expected to induce. A distributional Action-Wrench Critic is then trained on deployment rollouts to distinguish motions that look equally successful but end in different contact outcomes.
What the common framing gets wrong
The robot-foundation-model story of 2026 is told as vision-language-action: show the robot pixels and words, get behaviour. This result says that at assembly tolerances the visual channel is close to useless on its own — the strongest vision-led baseline lands at 15%. Two things must be kept straight before the 82% travels, though. It is measured after on-robot, part-specific adaptation of a lightweight bounded actor reusing the frozen representation, so it is not a zero-shot generalist figure. And it covers five tasks in one lab, unreplicated, in a preprint with no peer review.
The word “foundation” is doing a lot of work
The training corpus, ManuFacet-1K, is 1,000 hours of force-synchronised data across three embodiments and multiple manufacturing cells. That is a serious dataset for contact-rich manipulation and a rounding error next to the corpora the word “foundation” usually implies. The honest description is a specialist model for a physical regime, not a generalist.
Why the benchmark matters more than the winner
Sub-millimetre insertion is precisely the work humanoid and industrial-robot vendors keep promising factories and keep not delivering. Whatever one makes of Facet-0's own number, the 15% baseline is the useful datum: it is a published measure of how far vision-only policies sit from the tolerances real assembly requires.