[论文] LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

论文概要

研究领域: ML 作者: Zhixu Li, Keqian Tang, Litian Gong, Jingyu Yao, Wenqian Zhang, Tianze Xu, Zehao Wang, Yuping Wang, Junge Zhang, Jiachen Li 发布时间: 2026-10-15 arXiv: 2610.22372

中文摘要

个性化对于通用机器人进入家庭环境并实现真正的实用性至关重要。虽然通用机器人操作已被广泛研究，但个性化机器人操作（PeRM）缺乏一个能区分任务完成与偏好满足的控制基准。我们引入 LIBERO-PeRM，这是一个用于具身智能体偏好学习的基准，研究五个问题：1）策略如何在零样本设置下遵循偏好；2）如何从演示中学习单一偏好；3）随着演示增加，策略获取偏好的效率如何；4）学到的偏好如何跨布局和任务泛化；5）策略如何组合偏好并在冲突下表现。我们提出了模拟环境中可观察偏好的受控形式化和分类法，涵盖运动、技能和用户档案选择。基于 LIBERO，程序化流水线生成有偏好和无偏好的成对演示，创建了五个任务套件（312 个任务）。实验表明，当前 VLA 策略在零样本个性化方面表现不佳，而 SFT 使偏好可学习。混合训练比顺序自适应更好地保留无偏好性能，但可能削弱运动偏好学习；偏好泛化、分解和冲突处理仍然困难。

原文摘要

Personalization is essential for general-purpose robots to transition into household environments and achieve true practicality. While general robotic manipulation has been extensively studied, personalized robotic manipulation (PeRM) lacks a controlled benchmark for separating task completion from preference satisfaction. We introduce LIBERO-PeRM, a benchmark for preference learning in embodied agents. LIBERO-PeRM studies five questions: 1) how policies follow preferences in zero-shot settings; 2) how they learn single preferences from demonstrations; 3) how efficiently they acquire preferences as demonstrations increase; 4) how learned preferences generalize across layouts and tasks; and 5) how policies compose preferences and behave under conflicts. We propose a controlled formulation a...

#论文 #arXiv #ML #小凯