39 lines
1.2 KiB
Markdown
39 lines
1.2 KiB
Markdown
---
|
||
title: "Latent World Model (Robotics)"
|
||
created: 2026-06-24
|
||
updated: 2026-06-24
|
||
type: concept
|
||
tags: ["world-model", "jepa", "robot-learning", "latent-representation"]
|
||
sources:
|
||
- "[[vla-jepa-2026]]"
|
||
---
|
||
|
||
# Latent World Model (Embodied)
|
||
|
||
Latent World Model 是 VLA-JEPA 中的世界模型组件,基于 JEPA 范式在 latent space 中建模状态转移动态。
|
||
|
||
## 架构
|
||
|
||
- **Target Encoder**:V-JEPA2,frozen,从未来帧产生 latent world state targets
|
||
- **Predictor**:Autoregressive Transformer (12 层, 8 注意力头, 2048-dim)
|
||
- **注意力**:单时间步内双向(K 个 latent action token + N 个 image latent token),跨时间步因果
|
||
|
||
## 训练目标
|
||
|
||
$$\mathcal{L}_{WM} = \sum_{k=1}^{T} \mathbb{E}_{s_{t_k} \sim F(\cdot)} (\hat{s}_{t_k} - s_{t_k})$$
|
||
|
||
Target encoder F(·) 提供 ground-truth world state,predictor 学习预测。
|
||
|
||
可解释为 ELBO 最大化:
|
||
$$\log p(s_{1:T} | z_{0:T-1}) \geq \sum \mathbb{E}[\log p_\theta(\hat{s} | s)] - D_{KL}(F \| p_\theta^{WM})$$
|
||
|
||
## 与通用 World Model 的区别
|
||
|
||
不同于 Dreamer 等 pixel-space world model,Latent World Model 在语义空间运行,天然过滤像素噪声。
|
||
|
||
## 参考
|
||
- [[vla-jepa-2026]]
|
||
- [[jepa]]
|
||
- [[world-model-lecun]]
|
||
- [[leakage-free-state-prediction]]
|