TLDR: A new framework is introduced that uses Large Language Models (LLMs) to generate personalized explanations for human-robot interactions. By maintaining a dynamic ‘user knowledge-memory model’, the system adapts the level of detail in explanations, preventing cognitive overload. Experimental results show that a two-stage architecture, which first generates a full explanation and then personalizes it based on retrieved user knowledge, is most effective in reducing explanation detail when relevant user knowledge exists, demonstrating a significant advancement in explainable human-robot interaction.
In the evolving world of Human-Robot Interaction (HRI), a significant challenge lies in helping people truly understand how robots behave and make decisions. This area, known as eXplainable HRI (XHRI), focuses on creating explanations that are clear and effective. Previous research has highlighted that for these explanations to be truly useful and easy to grasp, they need to be personalized, adjusting the level of detail to suit each individual user.
This paper introduces a new framework designed to address this challenge. It allows robots to keep track of what a user already knows, building and updating a ‘knowledge-memory model’ over time. This memory then enables the robot to tailor its explanations, making them more concise when the user is already familiar with certain concepts, and more detailed when new information is being introduced.
The researchers evaluated three different architectural approaches within this framework, all utilizing Large Language Models (LLMs) to generate these personalized explanations. These architectures were tested in two practical scenarios: a robot patrolling a hospital and a kitchen assistant robot. The goal was to see which method was most effective at reducing the level of detail in explanations only when the user had relevant prior knowledge.
How the Framework Works
The core of the framework involves several key modules. A ‘persistent user memory’ stores a list of concepts the user has learned about the robot. A ‘knowledge extraction’ module populates this memory from previous conversations. When a user asks a question, a ‘retrieval’ module finds relevant concepts from this memory based on how similar they are to the current query. Finally, a ‘post-processing’ step adjusts these retrieved concepts, even accounting for the natural decline of memory over time using a forgetting curve, to create a ‘user context’ that guides the explanation’s detail level.
The three architectures differ mainly in how they query the user’s memory:
- Question Query: Uses the user’s direct question to retrieve related concepts.
- Robot Context Query: Uses information about the robot’s internal state and recent actions to retrieve memory.
- Explanation Query (Two-Stage): First generates a full, detailed explanation, and then uses this explanation itself as a query to personalize it further in a second stage.
Also Read:
- A New Framework for Personalized AI: Integrating Cognitive Memory and Thought Processes
- VerifyLLM: AI-Powered Verification for Robot Task Plans
Experimental Findings
The experiments, conducted using synthetic data across various topics for both robot use cases, revealed that the two-stage ‘explanation query’ architecture performed the best. This architecture was most effective at accurately retrieving relevant user knowledge and significantly reducing the length of explanations when the user already knew the topic. This is because the full explanation provides a richer, more semantically similar context for retrieving prior knowledge compared to just the user’s question or the robot’s internal context.
For instance, if a user had previously learned why the robot might mistake an orange for an apple due to bad lighting, and then later asked about a lemon being mistaken for a peach, the personalized explanation would be much shorter, referencing the previous, similar explanation. Crucially, the system also demonstrated a very low rate of false positives, meaning it didn’t incorrectly summarize explanations for topics the user had no prior knowledge about.
While the two-stage approach showed superior personalization, it does involve two LLM calls, which can potentially increase the time it takes to generate an explanation. However, if no relevant concepts are retrieved, the second stage is skipped, forwarding the initial full explanation, which can sometimes make it faster than other methods depending on how often users ask about known versus unknown topics.
This research marks a significant step towards creating more intuitive and user-friendly robots by making their explanations truly adaptive. Future work will involve real-world user studies to evaluate the practical impact of these personalized explanations on user experience and understanding. For more details, you can read the full paper here.
