As robots become more common in homes, one key challenge is teaching them not just to complete tasks, but to do so in ways that match individual user preferences. A newly published research paper introduces LIBERO-PeRM, a benchmark designed to test how well robots can personalize their actions based on user preferences, marking an important step toward making household robots more practical and user-friendly.
Key Takeaways
- LIBERO-PeRM is the first benchmark focused specifically on measuring how robots learn and follow personal preferences separate from just completing tasks.
- The benchmark covers five key questions, including how robots perform without prior preference knowledge, how they learn from demonstrations, and how they handle conflicting preferences.
- Experiments show current robot learning policies struggle with zero-shot personalization but improve when trained with preference examples.
- While some training methods help maintain baseline task performance, challenges remain in generalizing preferences and dealing with conflicting user demands.
Robots today can often perform a wide range of physical tasks, from picking up objects to assembling items. However, these robots typically focus on the task itself—like moving a cup from one place to another—without considering individual user preferences such as how quickly the robot moves, what grip style it uses, or other subtle choices that make the interaction feel more natural and tailored. The LIBERO-PeRM benchmark introduced by Zhixu Li, Keqian Tang, Litian Gong, and colleagues addresses this gap by providing a controlled environment to study how robots can learn and apply these preferences.
The researchers developed LIBERO-PeRM using simulation, where robots interact with virtual household environments. This approach allows them to isolate and test specific preferences expressed in natural language—like preferring gentle motions or specific skill styles—without changing the fundamental task goals. The benchmark includes five task suites comprising 312 tasks, each with paired demonstrations showing how to complete tasks with and without preferences. This setup helps evaluate not just if the robot can finish a job, but how well it aligns with the desired personal touch.
To explore how robots learn preferences, the study examined multiple scenarios: zero-shot settings where the robot has no prior examples, learning from a single demonstration, gradually improving as more demonstrations are provided, generalizing learned preferences to new tasks or layouts, and handling situations where preferences might conflict (for example, wanting both speed and gentleness). They tested several training methods, including supervised fine-tuning (SFT) and mixed training approaches that combine preference and no-preference data.
The results showed that current vision-language-action (VLA) policies often struggle to personalize actions without prior examples. However, when trained with preference demonstrations, robots became better at adapting their behavior. Mixed training methods helped maintain the ability to perform tasks even when no preferences were given, though some trade-offs appeared, such as weaker learning of motion-related preferences. Generalizing preferences to new contexts and resolving conflicts between competing preferences remained challenging areas.
This research highlights the complexity of making robots truly personalized assistants. By providing a structured benchmark and dataset, LIBERO-PeRM offers a valuable tool for the robotics community to develop and test new algorithms that go beyond task completion to satisfy individual user needs. In the future, advances built on this work could lead to household robots that not only get the job done but do so in ways that feel intuitive and customized for each person.
Based on research published on arXiv by Zhixu Li, Keqian Tang, Litian Gong et al..