N₀-TWAM trains robots to predict touch, and it lifts contact-task success to 46.3%

imjustnewatai · x · 2026-07-29

N₀-TWAM teaches robots to predict touch, not just vision

The paper argues that cameras alone often cannot tell whether a grasp, plug-in, or insertion actually succeeded. N₀-TWAM therefore predicts three things together: the next camera view, expected fingertip contact, and the action likely to produce both, while also reading current force signals.

What it was trained on

A 7.16B-parameter model
About 7.5 million clips
Data from 6 robot bodies and 450 tasks
Each clip aligns video, language, actions, and per-finger touch

Results

46.3% average success across 8 real contact-heavy tasks
Best vision-language-action baseline: 30%
On bottle standing, it reached 70% vs 20% for the best baseline
On mostly visual placement tasks, vision-only models matched or beat it

Why the author thinks it matters

Removing predicted future touch dropped the benchmark mean from 67.0% to 56.4%
Removing current touch dropped it further to 50.0%
The main gains seem to come exactly where cameras miss physical state

The post frames touch as a verifier for physical AI: a robot should not advance unless force/contact confirms that the action really happened. The author also points to a possible data moat, since synchronized tactile data is rare compared with ordinary internet video.

Original post →

N₀-TWAM trains robots to predict touch, and it lifts contact-task success to 46.3%

More from Embodied