High-Level Planning
把'去厨房热一杯牛奶'拆成一长串可执行步骤——LLM 当大脑，机器人当手脚。这一族研究关心：怎么让大模型只输出可执行命令，并对失败做出反应。
先读这三篇。
SayCan 教 LLM 看清自己能做什么 → Code-as-Policies 让它直接写代码 → Inner Monologue 让它反思失败。
-
1
SayCan: Do As I Can, Not As I Say
让"见多识广但出不了门的 AI"出主意，让机器人自己摸口袋说"这事我现在能做"，两边都点头才动手。
-
2
Code as Policies: Language Model Programs for Embodied Control
你说一句"把方块叠进碗里"，AI 当场写几行 Python 代码，机器人立刻照着跑。不用提前教它新动作。
-
3
Inner Monologue: Embodied Reasoning through Planning with Language Models
让机器人边干活边在心里念叨：看到啥、做成没、人改主意没，全翻成文字塞回 AI，它就能边做边改计划。
2022 到 2024，11 篇怎么排开。
祖师爷
经典
前沿
High-Level Planning 全部 11 篇。
| era | year | title | venue |
|---|---|---|---|
| 祖师爷 | 2022 | SayCan: Do As I Can, Not As I Say | CoRL |
| 祖师爷 | 2022 | Inner Monologue: Embodied Reasoning through Planning with Language Models | CoRL |
| 祖师爷 | 2023 | Code as Policies: Language Model Programs for Embodied Control | ICRA |
| 祖师爷 | 2023 | LLM+P: Empowering LLMs with Optimal Planning | arXiv |
| 祖师爷 | 2023 | PaLM-E: An Embodied Multimodal Language Model | ICML |
| 祖师爷 | 2023 | ProgPrompt | ICRA |
| 经典 | 2023 | ChatGPT for Robotics | IEEE Access |
| 经典 | 2023 | VoxPoser | CoRL |
| 经典 | 2024 | GenSim | ICLR |
| 经典 | 2024 | RoboFlamingo | ICLR |
| 经典 | 2024 | Tree-Planner | ICLR |