LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

TL;DR

LIBERO-PeRM introduces a simulation-based benchmark for personalized robotic manipulation, enabling controlled evaluation of how agents learn and apply user preferences. It features 312 tasks with paired demonstrations, structured preference categories, and automated metrics to assess preference adherence across zero-shot, learned, and conflicting scenarios.

Why It Matters

Personalized robotic manipulation is critical for real-world deployment in homes, but lacks standardized evaluation. LIBERO-PeRM provides a reproducible framework to assess how well robots adapt to individual user preferences, advancing the development of user-centered robotic systems.

Key Findings

• Current vision-language action (VLA) policies fail to apply preferences in zero-shot settings without prior training.
• Supervised fine-tuning enables learning of single preferences, and mixed training improves performance with contextual prompts while preserving baseline task performance.
• Generalization of preferences across layouts and tasks, decomposition of composite preferences, and resolution of preference conflicts remain significant challenges.