Robot Behavior Personalization from Sparse User Feedback

TL;DR

The paper tackles personalizing robot assistance for an open, unbounded set of household tasks using sparse user feedback. It introduces TAACo, a two-module framework comprising a Commonsense Module that maps task components to abstract concepts and a Personalization Module that predicts the preferred adaptation $\phi$ and generates explanations, with $\tilde{t}=\{(x,\theta^x,m)\}$ serving as the intermediate representation. TAACo demonstrates strong empirical performance, achieving $\approx 0.71$ prediction accuracy with 40 user-feedback samples (oracle $=0.89$) and explaining its decisions effectively, outperforming GPT-4 and a rule-based baseline, while enabling per-user local training and explainability. The approach is validated on real-user data from five older adults and demonstrated on a Stretch robot, highlighting its potential to deliver personalized, interpretable, and data-efficient home-robot assistance. Overall, TAACo advances open-set personalisation in HRI by leveraging abstract concepts to bridge user preferences with robot policies and by producing faithful explanations that align with user reasoning.

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 customize 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.

Figures (6)

• Figure 1: Task Adaptation using Abstract Concepts (TAACo) learns user preferences regarding how they want the robot to assist with an open set of household tasks from limited user feedback. This requires commonsense reasoning to extract relevant semantic information regarding a novel task, and personalization based on limited feedback. In addition, TAACo can explain its predictions to the user in an intuitive manner.
• Figure 2: TAACo generates a concept-based representation for a given task through a Commonsense Module, and uses it to predict the preferred action adaptation through a Personalization Module.
• Figure 3: Comparison of our model against GPT and rule-based baselines, as well as against using an oracle version of commonsense, on (a) overall prediction accuracy over varying amounts of user feedback, and (b) a breakdown of the errors, with 40 feedback samples, with the each bars representing the total error rate and each color representing an error type, along with an example in a box of the same color
• Figure 4: Comparison of Explanation Accuracy against baselines.
• Figure 5: Ablation results of our model, removing explanation training, and removing the use of concepts entirely.