GPT-6 Astra edits robot embeddings to fix paraphrased instructions
A Stanford study by Jusheng Zhang, Keze Wang, and Liang Lin used GPT-6 Astra to edit 65 embedding rows in an OpenPI π₀.₅ LIBERO vision-language-action policy. In simulation, success on 335 held-out instructions containing modified tokens rose from 74.3% to 87.5%, while outcomes on 430 other episodes remained unchanged; the full test also included four regressions. The editing used practice-sample feedback and tool calls without gradient updates or additional robot demonstrations, and the paper, code, weights, and experiment records were released.
The report presents an open, auditable method for using an LLM to make a robot policy more robust to alternative wordings, though the main evaluation used one random seed in simulation and has not established real-world transfer.
机器之心 2026-11-13 14:28 北京
含被改词元的 335 条留出指令,成功率从 74.3%升至 87.5%;另外 430 个回合,成败结果逐一不变。
只改 65 行词嵌入,GPT-6 Astra 修复了一类机器人「换种说法就失败」的问题。
任务是打开炉灶。研究者给机器人的指令是「activate the cooking appliances heat source」——启动烹饪设备的热源。原始策略却伸手去拿盘子,执行 300 步后仍未完成任务。
#GPT-6 Astra#OpenPI π₀.₅ LIBERO#LIBERO#LIBERO-Plus#Stanford University#OpenPI#FastWAM