Problem
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.
Method
We introduce LIBERO-PeRM, a benchmark for preference learning in embodied agents. We propose a controlled formulation and taxonomy of observable preferences in simulation, covering motion, skill, and user-profile choices. This formulation is restricted to language-expressible preferences that can be instantiated in LIBERO without changing the task goal, evaluated automatically, and reused across tasks. Building on LIBERO, our procedural pipeline generates paired demonstrations with and without preferences. We create five task suites (312 tasks) with high-quality demonstrations, rich metadata, and preference-satisfaction metrics.
Findings
Experiments show that current VLA policies struggle with zero-shot personalization, while SFT makes preferences learnable. Mixed training achieves better no-preference retention than sequential adaptation and improves performance with contextual prompts compared with preference-only training, but can weaken motion-preference learning; preference generalization, decomposition, and conflict handling remain difficult.
Limitations
The benchmark is limited to simulation and language-expressible preferences. The study does not cover all possible preference types or real-world robotics.
Practical Significance
LIBERO-PeRM provides a controlled benchmark for preference learning in embodied agents, enabling systematic study of how robots can learn and adapt to user preferences, which is crucial for household robotics.