LIBERO-PeRM test for home robots: completing tasks does not guarantee meeting user preferences

It is not enough for a robot to succeed at grasping a cup or cleaning a table to be ready to live inside homes; the gap between completing a bare task and performing it according to the homeowner’s preference has remained absent from rigorous scientific measurement. This technical challenge prompted a research team to introduce the LIBERO-PeRM benchmark in the proceedings of the NIPS 2026 conference, aimingDecoupling task completion measurement from user preference satisfactionwhich opened the door to the first systematic evaluation of embodied agents’ ability to incorporate behavioral customization.

The researchers built their benchmark on the LIBERO simulation environment, developing an operational pipeline that generates demonstration trajectories paired with tasks and their associated preferences. In total, the team produced 312 tasks distributed across five comprehensive test suites. They classified the observed preferences into three clear categories: motion options, skill choices, and user profile preferences, limiting them to linguistically expressible preferences that can be applied without altering the task’s final goal. This enables precise, automated measurement of user satisfaction and the reuse of preferences across multiple tasks.

In the test details we followed five critical research questions posed by the paper: how policies implement preferences in the first experimental environment without prior training, how they learn a single preference from demonstrations, how their efficiency improves as those demonstrations increase, how well what they have learned generalizes to new environments and tasks, and finally how they combine multiple preferences and act when those preferences conflict. Experimental results revealed that current vision-language-action (VLA) policies suffera noticeable inability to customize without prior training, as they fail to satisfy user wishes merely upon receiving a linguistic command, confirming that customization does not arise automatically from foundational models.

Experiments demonstrated that supervised fine-tuning makes user preferences learnable within models, yet the training approach imposes precise technical trade-offs; mixed-training, which combines preference-free data with curated data, achieved higher retention of core tasks compared with sequential adaptation, which degrades original performance. Moreover, mixed-training showed superiority with contextual commands over training on preferences alone, but this advantage was accompanied by a genuine decline in the accuracy of learning the same motion preferences, while challenges in generalizing preferences, disentangling them, and resolving conflicts among them remain unsolved by current methods.

This distinction between task completion and operator preference reshapes the priorities of engineers and development teams in the Gulf and Egypt, especially as hospitality projects, home automation, and smart warehouses accelerate across the region. Purchasing a robot or embodied model that can maneuver mechanical arms is no longer sufficient to build a service acceptable to the end user; the difference lies in the skill of constructing demonstration pairs and training models with mixed-training that preserves core device abilities without destroying their responsiveness to local environmental specifics. This shift means thatthe actual project cost will shift from hardware acquisition to balancing data packages and fine-tuningto avoid robot behavior fragmentation when operational preferences conflict.

NIPS 2026 research reveals that embodied intelligence does not face a crisis in understanding what to do as much as it faces a crisis in understanding how a human wants it done. Confronted with these technical limits, your practical step as an engineer or technical lead in regional automation projects becomes verifying the combined fine-tuning mechanisms and model behavior when instructions conflict, rather than assuming that large models will automatically capture home or field work patterns simply by being given verbal guidance.