Personalised Explanations in Long-term Human-Robot Interactions

Ferran Gebellí, Anaís Garrell, Jan-Gerrit Habekost, Séverin Lemaignan, Stefan Wermter, Raquel Ros

TL;DR

This paper introduces a framework for robots to provide personalized explanations by "remembering" what a user already knows. By integrating a persistent memory model with Large Language Models (LLMs) and an Ebbinghaus-inspired forgetting curve, the system can dynamically shorten explanations for familiar concepts while providing full detail for new ones, preventing cognitive overload in long-term human-robot interactions.

What Problem Does This Paper Address?

In the field of eXplainable Human-Robot Interaction (XHRI), robots need to explain their decisions (e.g., "Why didn't you pick up that apple?"). However, most current systems provide the same level of detail every time. In long-term scenarios—like a robot assistant in a hospital or kitchen—this leads to redundant, overly wordy explanations that can frustrate or confuse users.

The core challenge is personalization: a robot should estimate the user's "mental model" and adjust its level of detail accordingly. If a robot has already explained that it struggles with bad lighting, it shouldn't repeat the full technical lecture a week later; it should simply reference the previous explanation.

Methods

The authors propose a framework consisting of four key modules:

  1. Persistent User Memory: A dynamic list of concepts (e.g., "The robot can be obstructed by obstacles") stored as text embeddings.
  2. Knowledge Extraction: A dedicated LLM that "listens" to the explanations the robot gives and extracts key facts to update the user's memory.
  3. Retrieval of Relevant Knowledge: When a user asks a question, the system uses Retrieval-Augmented Generation (RAG) with a similarity threshold ($\tau_{retr}$) to find related concepts in the user's memory.
  4. Post-processing (The Forgetting Curve): To make the model realistic, the authors apply the Ebbinghaus forgetting curve ($P = e^{-\lambda t}$). This calculates the probability that a user still remembers a concept based on how much time has passed, allowing the robot to provide more detail if the information was shared a long time ago.

The researchers compared three architectural variants for the retrieval query:

  • Question Query: Uses the user's raw question to search memory.
  • Robot Context Query: Uses the robot's internal logs to search memory.
  • Explanation Query (Two-Stage): First generates a full, detailed explanation, then uses that explanation to search memory for overlaps, and finally adapts the text into a personalized version.

Datasets

The study used synthetic data to maintain controlled interactions for systematic analysis. They developed two distinct use cases:

  • Kitchen Assistant (NICOL robot): 7 topics involving grasping failures, unreachable objects, and perception errors.
  • Hospital Patrolling (Tiago LITE robot): 7 topics involving alarms, obstructed paths, and hardware issues (e.g., motor overheating).

For each topic, they created a pool of questions and hypothetical events to simulate interactions over different time scales (1 week vs. 1 month).

Key Insights and Ideas

  • The Two-Stage Approach Wins: The "Explanation Query" architecture performed best. By generating the full explanation first and using it as the search query, the system found much more relevant matches in the user's memory compared to just using the user's question.
  • Semantic Similarity is Key: User questions are often vague (e.g., "Why didn't you pick that up?"). The robot's internal state is technical. The explanation itself sits in the middle, providing the best semantic bridge to retrieve past knowledge.
  • Thresholding for "Unknowns": By using a similarity threshold rather than just picking the "top K" results, the system successfully avoided giving "summarized" (short) explanations for topics the user actually didn't know yet.
  • Modeling Forgetting: Integrating a decay parameter for memory allows the robot to be "patient." If a concept hasn't been discussed in a month, the robot automatically provides more detail, assuming the user might have forgotten.

Key Results

  • High Precision: The two-stage architecture achieved a retrieval precision of 94.27% to 96.15% for known topics.
  • Effective Length Reduction: For topics the user recently learned (1 week ago), the system reduced explanation length by approximately 39%.
  • Robustness: For entirely new topics, the system correctly identified that the user had no prior knowledge, resulting in a negligible length reduction (only ~2.5%), ensuring the user still received the full necessary information.
  • Inference Trade-off: While the two-stage approach is more accurate, it requires two LLM calls, increasing the token count and processing time (averaging 3.88 seconds per explanation).

Limitations and Future Work

  • Synthetic Evaluation: The current study relies on synthetic data and LLM-based evaluation. The authors acknowledge the need for a longitudinal user study with real humans to measure actual usability and trust.
  • Inference Latency: The two-stage process is slower. Future work could explore optimizing these calls or using smaller, specialized models for the extraction/adaptation phases.
  • Multimodal Memory: The current memory is purely text-based. The authors suggest future versions could populate memory from other sources, like what the user has physically seen the robot do (observational learning).

Summary updated

Abstract

In the field of Human-Robot Interaction (HRI), a fundamental challenge is to facilitate human understanding of robots. The emerging domain of eXplainable HRI (XHRI) investigates methods to generate explanations and evaluate their impact on human-robot interactions. Previous works have highlighted the need to personalise the level of detail of these explanations to enhance usability and comprehension. Our paper presents a framework designed to update and retrieve user knowledge-memory models, allowing for adapting the explanations' level of detail while referencing previously acquired concepts. Three architectures based on our proposed framework that use Large Language Models (LLMs) are evaluated in two distinct scenarios: a hospital patrolling robot and a kitchen assistant robot. Experimental results demonstrate that a two-stage architecture, which first generates an explanation and then personalises it, is the framework architecture that effectively reduces the level of detail only when there is related user knowledge.
BibTeX
@inproceedings{gebelli2025personalised,
  author = {Ferran Gebellí and Anaís Garrell and Jan-Gerrit Habekost and Séverin Lemaignan and Stefan Wermter and Raquel Ros},
  title = {{Personalised Explanations in Long-term Human-Robot Interactions}},
  booktitle = {RO-MAN},
  year = {2025},
  doi = {10.1109/ro-man63969.2025.11217900},
}

Loading related content...

Figure