Stanford Fixes Robot Instruction Failures by Editing 65 Embedding Rows

A Stanford-led team proposes debugging robot policies by directly editing 65 rows of a vision-language model's token embedding layer, with no gradient updates or new demonstration data. On 335 held-out instructions containing the edited tokens, success rate rose from 74.3% to 87.5%, while outcomes on 430 other episodes stayed unchanged. The paper, code and all 16 experiment logs are public.