Current robot learning systems struggle to personalize behavior to user preferences without explicit training, and even with training, they have difficulty composing multiple preferences or generalizing them across different tasks and environments.
LIBERO-PeRM is a benchmark for teaching robots to follow user preferences in manipulation tasks. It includes 312 simulated tasks with demonstrations showing how to complete actions with different style preferences (like moving slowly vs. quickly), and tests whether robots can learn these preferences from examples, generalize them to new situations, and handle conflicting preferences.