Robot Behavior Personalization from Sparse User Feedback

Maithili Patel, Sonia Chernova

Workshop Statement:

The rise of large language models have introduced semantic knowledge at a scale that was previously unreachable. The scale of these models helps them generalize to a wide variety of tasks and objects through general commonsense knowledge, but creates a challenge of adapting to unique preferences of individuals. It is intractable to obtain foundational-model-level data to finetune them into bespoke models. In this work, we take an alternative approach, by utilizing the commonsense knowledge in LLMs to create a rich intermediate representation, which can then be used by a custom Personalization module to make decisions specific to the individual.

We contribute Task Adaptation using Abstract Concepts (TAACo), which addresses the challenge of personalizing robot behavior to the user’s preferences of how a robot should assist the user with a specific task. We formulate this as predicting for each task, an adaptation out of doing a task now, doing it later, reminding the user about it or not doing anything about it. We learn such preferences from small amounts of user feedback on other tasks in the past.

The key idea of TAACo is to utilize abstract concepts related to the task as my intermediate representation. For instance, the fact that a vase is *fragile* can indicate that the robot should not do a task, or the *mundaneness* of a task can indicate that the robot should do it. The use of abstract concepts explicitly has a few advantages: 1) knowledge of how tasks relate to abstract concepts is relatively universal and can be extracted from LLMs, 2) humans tend to reason using such abstract concepts and provide feedback in that manner as well, which can be used in model training, 3) the smaller set of abstract concepts, relative to the set of all tasks, allows a small personalized model to generalize to an open-set of tasks, 4) the model can explain its decisions in the form of abstract concepts. Moreover, the flexibility of transformers used in our Personalization module, and the open-set knowledge of LLMs, allows TAACo to be flexible to an open set and changing number of concepts. Ultimately, we train my Personalization model to use the rich intermediate representation to determine what to do about a task. We train it not only on the output predictions, but also on abstract concepts in user-provided feedback, by explicitly training the attention weights. Such alignment provides further scaffolding to the model's training, avoiding spurious correlations and enabling efficient training.

In all, the combination of commonsense from LLMs and a smaller personalization model, allows TAACo to both **generalize** to a an open-set of tasks, and **personalize** to user preferences, given limited feedback from the user.

Keywords: Human-Centered Robotics, Domestic Robotics, Personalization

TL;DR: We contribute TAACo, which adapts robot behavior to the user’s preferences of how a robot should assist the user with a specific task over an open-set of household tasks from limited user feedback.

Abstract: As service robots become more general-purpose, they will need to adapt to their users' preferences over a large set of all possible tasks that they can perform. This includes preferences regarding which actions the users prefer to delegate to robots as opposed to doing themselves. Existing personalization approaches require task-specific data for each user. To handle diversity across all household tasks and users, and nuances in user preferences across tasks, we propose to learn a task adaptation function independently, which can be used in tandem with any universal robot policy to personalize robot behavior. We create Task Adaptation using Abstract Concepts (TAACo) framework. TAACo can learn to predict the user's preferred manner of assistance with any given task, by mediating reasoning through a representation composed of abstract concepts built based on user feedback. TAACo can generalize to an open set of household tasks from small amount of user feedback and explain its inferences through intuitive concepts. We evaluate our model on a dataset we collected of 5 people's preferences, and show that TAACo outperforms GPT-4 by 16% and a rule-based system by 54%, on prediction accuracy, with 40 samples of user feedback.

Submission Number: 16