GPT-6 Astra explores the scalable growth of embodied intelligence: incorporating human action demonstrations to enhance model reliability

GPT-6 Astra demonstrates zero-shot capability on operational tasks, but the reliability of strategies in actual deployment still faces challenges. Experiments with 96 real devices show that human action demonstrations can effectively improve the model's performance in interaction, alignment, and long-process tasks. This research provides new insights for the scalable growth of embodied intelligence.