LIBERO-PeRM

LIBERO-PeRM is the first controlled benchmark for preference learning in embodied agents, specifically evaluating task completion and preference satisfaction separately.

LIBERO-PeRM divides observable preferences in simulation into three categories: motion preferences, skill preferences, and user profile preferences.

LIBERO-PeRM imposes four hard constraints on preference definition: preferences must be expressible in language, must not change the task goal, must be automatically evaluable, and must be reusable across tasks.

LIBERO-PeRM designs its evaluation system around five progressive questions: zero-shot preference following, single-preference learning, preference acquisition efficiency, preference generalization, and multi-preference composition and conflict handling.

LIBERO-PeRM builds a procedural data generation pipeline on top of the LIBERO environment, ultimately constructing five task suites containing 312 tasks, with high-quality demonstrations, rich metadata, and preference satisfaction metrics.

The paper's experiments show that current mainstream VLA policies perform poorly in zero-shot personalized scenarios: they can complete tasks, but it is difficult for them to follow preferences from instructions alone.

The paper's experiments show that supervised fine-tuning (SFT) can make preferences learnable.

The paper's experiments show that mixed training outperforms sequential adaptation on preference-free retention, and also performs better than pure preference training when combined with in-context prompting, but it weakens the learning of motion preferences.

The paper explicitly points out that generalization, decomposition, and conflict handling of preferences remain difficult, and the preferences models learn are often bound to specific scenarios, failing easily when the layout is changed or preferences are stacked.

LIBERO-PeRM's code, dataset, and demonstrations have been made public on the project website.