LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

Forum topic · 小凯 · 2026-10-17

Summary

LIBERO-PeRM is a new benchmark for personalized robotic manipulation (PeRM) that separates task completion from preference satisfaction in embodied agents. Built on LIBERO, it uses a programmatic pipeline to generate paired preference-aware and preference-free demonstrations across five task suites totaling 312 tasks. The benchmark evaluates five questions: zero-shot preference following, learning a single preference from demonstrations, sample efficiency as demonstrations increase, generalization across layouts and tasks, and preference composition under conflicts. It introduces a controlled formalization and taxonomy of observable preferences in simulation covering motion, skills, and user profile selection. Experiments show current vision-language-action (VLA) policies perform poorly at zero-shot personalization, while supervised fine-tuning makes preferences learnable. Mixed training better preserves preference-free performance than sequential adaptation, but may weaken motion preference learning; preference generalization, decomposition, and conflict handling remain open challenges. Paper: arXiv 2610.22372.

Paper Overview

Field: Machine Learning Authors: Zhixu Li, Keqian Tang, Litian Gong, Jingyu Yao, Wenqian Zhang, Tianze Xu, Zehao Wang, Yuping Wang, Junge Zhang, Jiachen Li Published: 2026-10-15 arXiv: 2610.22372

Abstract

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. The authors 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 5. How policies compose preferences and behave under conflicts

Key Contributions

A controlled formulation and taxonomy of observable preferences in simulation, covering motion, skills, and user profile selection
A programmatic pipeline built on LIBERO that generates paired preference-following and preference-free demonstrations
Five task suites comprising 312 tasks

Findings

Current VLA (vision-language-action) policies perform poorly on zero-shot personalization
SFT (supervised fine-tuning) makes preferences learnable
Mixed training preserves preference-free performance better than sequential adaptation, but may weaken motion preference learning
Preference generalization, decomposition, and conflict handling remain difficult open problems

--- *Auto-collected on 2026-10-17.*