La-JEPA 用潜空间世界建模训练紧凑机器人策略
"La-JEPA: Compact Robot Policies with Latent World Modeling"
摘要
alphaXiv 介绍论文 La-JEPA，它用预训练的 V-JEPA 视觉表征训练紧凑机器人策略，训练时同时预测未来场景表征和机器人动作，推理时只生成动作。在 RoboTwin 2.0 上，其 1B 模型达到 60.8%，对比 12.4B 的 Fast-WAM 为 39.9%，运行速度快 5 倍，并在四个任务上取得与 π0.5 相当的真实机器人表现。
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
"La-JEPA: Compact Robot Policies with Latent World Modeling"
What if robots could outperform massive video-generation models with a fraction of the parameters, simply by learning to predict the future without actually generating it?
This paper La-JEPA trains a compact robot policy using pretrained V-JEPA visual representations. During training, it learns to predict both future scene representations and robot actions, but at inference, it only generates actions.
On RoboTwin 2.0, its 1B model achieves 60.8%, compared to 39.9% on 12.4B Fast-WAM. It also runs 5x faster and achieves comparable real-robot performance to π0.5 across four tasks.
来源：alphaXiv · x.com
引用原文
askalphaxiv. "La-JEPA: Compact Robot Policies with Latent World Modeling"[EB/OL]. (2026-11-10)[访问日期]. https://x.com/askalphaxiv/status/2120062613589840261.