Imitation Learning
机器人怎么学技能?最朴素的答案:抄作业。研究员用遥操作设备演示,机器人从轨迹里学。这一族关心:演示怎么采、怎么编码、怎么少演示几次也能学会。
先读这三篇。
DAgger 解决误差累积 → ACT/ALOHA 把遥操作做成标配 → UMI 用手持夹爪+GoPro 采野外数据。
-
1
A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning
光看老师开车的录像不够 — 学生一走偏就越错越离谱。DAgger 让学生自己先开几圈,把走偏的地方拿去问老师答案,再训,反复几轮就稳了。
-
2
Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT/ALOHA)
几千美元搭一套双臂遥控器(ALOHA)让人录 50 次示范,机器人就学会一段一段动(ACT),能完成穿扎带这种细活。
-
3
Universal Manipulation Interface
人手拿一个"带摄像头的夹子"在厨房自己做事,录下来就能教机器人,全程不用机器人在场。
2011 到 2025,15 篇怎么排开。
祖师爷
经典
前沿
Imitation Learning 全部 15 篇。
| era | year | title | venue |
|---|---|---|---|
| 祖师爷 | 2011 | A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning | AISTATS |
| 祖师爷 | 2016 | Generative Adversarial Imitation Learning | NeurIPS |
| 经典 | 2021 | Implicit Behavioral Cloning | CoRL |
| 经典 | 2022 | Behavior Transformers: Cloning k Modes with One Stone | NeurIPS |
| 经典 | 2023 | Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT/ALOHA) | RSS |
| 经典 | 2023 | AnyTeleop | CoRL |
| 经典 | 2023 | RoboCat | TMLR |
| 前沿 | 2024 | ALOHA 2 | Tech Report |
| 前沿 | 2024 | DexCap | RSS |
| 前沿 | 2024 | HumanPlus | CoRL |
| 前沿 | 2024 | Mobile ALOHA | CoRL |
| 前沿 | 2024 | Universal Manipulation Interface | RSS |
| 前沿 | 2024 | Behavior Generation with Latent Actions (VQ-BeT) | ICML |
| 前沿 | 2025 | Generalizable Humanoid Manipulation with 3D Diffusion Policies (iDP3) | RSS |
| 前沿 | 2025 | SmolVLA | arXiv |