Robot Policy Metadata Conditioning Fails: Text Prompt Completely Ignored

DominiqueCAPaul · x · 2026-09-30

The author tried to replicate physicalint's approach of adding metadata about the episode or action chunk to the text prompt for the π0.7 model, using speed as the conditioning variable: the dataset contained two recording sessions, one with double the speed of the other, and a suffix "speed: fast|slow" was appended to the prompt.

Result: the policy completely ignored the text prompt, and behavior did not change when switching between "slow" and "fast".

Further analysis showed the speed information could already be extracted from the images themselves — a logistic regression distinguished the two sessions on 99.8% of held-out frames. The only visible difference was a slightly different placement of the box and tray on the table, but that was apparently enough.

Takeaway: if the conditioning information is already implicit in the observation images, extra metadata conditioning may be ignored by the policy; check for such leakage before running these experiments.

Original post →

More from Embodied

Embodied channel →