Search
Research explores the feasibility of using large models to debug robot policy weights. In the experiment, GPT-6 Astra—without gradient updates or additional demonstration data—modified only 65 lines of word embedding vectors in a vision-language-action robot model, raising the success rate on 335 held-out instruction tasks containing the modified tokens from 74.3% to 87.5%, while the per-episode success or failure outcomes for the other 430 unrelated instructions remained completely unchanged. Controlled experiments show that the improvement stems from the reasonable vocabulary embedding mappings discovered by Astra. All experimental data, code, and modification records have been made public. The current results are all from a simulation environment, and real-robot validation is still underway.
-
C114 Communication Network -
Communication Home
