How to?

papers batch 5

Jun 30, 2026
papers
1 Minutes
187 Words

Behavior Prompting Policy: Demonstrations as Prompts for Manipulation (35)

⭐️⭐️⭐️ 将一次性演示作为 in-context 任务提示 [Stanford, Austin Patel, Shuran Song]

🦾 https://behavior-prompting.github.io📃 https://hjfy.top/arxiv/2606.30457https://www.alphaxiv.org/abs/2606.30457💻 https://github.com/real-stanford/behavior_prompting

recap: diffusion policy 使用 obs_encoder(obs) 获得 global_cond 并直接 FiLM 调制 UNet 中的多层 x.

本文仅需在上述 obs_encoder 内部 cross-attn 一条演示下采样得到的 [enc(rgb), mlp(propio), mlp(action)],实现了 1. infer-time 指定折叠方式(训练时见过多种折叠方式); 2. infer-time 指定绘制数字(训练时未见过的数字)。实验发现需要训练数据多样性高时才能很好地 steering,当然,还是无法跨 action primitive 的.

这篇文章 abs 比我们写的好多了:

  1. We study Bahavior Propmting.
  2. To enable this, we present contributions in algo, data and evaluation.
  3. For algo, we.. For data we identify that .. is the primer driver and introduce iPhUMI.
Article title:papers batch 5
Article author:Julyfun
Release time:Jun 30, 2026
Copyright 2026
Sitemap