pi0: A Vision-Language-Action Flow Model for General Robot Control
Overview
pi0基本确立了之后VLA的几个重要的内容：
- pre-training + post training两个阶段训练的scheme
- 确立action expert的flow matching/diffusion process for continuous action outputs.
- 论文证明pre-training有效果
- 论文也论证了引入VLM从而有的“language instructions following”的特性，这也应该是为之后的subtask的生成做准备。
Pipeline
具体的attention mask为：
| Prefix | State | Action | |
|---|---|---|---|
| Prefix | 1 | 0 | 0 |
| State | 1 | 1 | 0 |
| Action | 1 | 1 | 1 |