Who Is the Agent Talking To? Participant-Aware Role Switching for Embodied Student-Support Dialogue
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Who Is the Agent Talking To? Participant-Aware Role Switching for Embodied Student-Support Dialogue".
Dev: Here's the gist of the paper. RPO-PDT presents a retrieval-grounded, role-play-based dialogue system for adaptive student support in higher education.
Rosa: RPO-PDT is able to provide institution-specific Personal Development Tutor (PDT) guidance using structured knowledge sources and is designed around a reverse-roleplay loop where unresolved interactions are replayed from the student perspective, enabling alternative tutor strategies to be generated and stored as reusable strategy memory.
Dev: This system supports both textbased and Furhat-based embodied interaction for demonstrating grounded, safe, and adaptive student-support dialogue.
Rosa: System Architecture and Components RPO-PDT is implemented using a modular architecture that combines Rasa-based dialogue orchestration, structured institutional knowledge sources, policy-constrained LLM generation, deterministic safety escalation, and an optional Furhat robot interface.
Title and authors: Rosa: They start by introducing RPO-PDT and how it provides institution-specific Personal Development Tutor guidance using structured knowledge sources. This grounding is key for making sure the advice is locally accurate.
Dev: The paper mentions they use YAML files as lightweight stores for all that knowledge, including course lookups and university support service information.
Rosa: And they also stress that the system is constrained by explicit persona, boundary, confidentiality, and safety policies embedded right into it.
Taro: That constraint mechanism sounds important because it’s not just letting an LLM talk freely; there are hard rules governing what the agent can actually do or say.
Dev: And they detail a deterministic safety escalation layer that checks both the user input and the generated response for risk levels like low, high, or crisis.
Rosa: The paper shows this safety layer is always active, checking for things like crisis disclosures and signposting to resources instead of just relying on generative modeling.
Taro: I wonder how robust that deterministic check is when dealing with complex human language; does it catch everything or just the obvious crises?
Dev: They say the purpose of that guideline isn't to make clinical judgments but to promote conservative support behaviors, meaning it steers the agent away from things like making medical diagnoses.
Rosa: It’s about ensuring that even when the system is trying to be helpful, it stays within safe operational boundaries defined by those YAML stores.
Taro: So for someone who only listens to this show, what does that safety layer actually mean in practice for the student interaction?
The paper's summary: Dev: The core concept they summarize is RPO-PDT’s ability to provide grounded support while actively learning from past interactions.
Rosa: They focus on the reverse-roleplay mechanism again, explaining that if a case isn't resolved, it gets stored in a strategy memory.
Dev: When we replay that unresolved case from the student's perspective, the system generates alternative tutor responses and saves them as reusable strategies for future use.
Taro: So it’s not just about solving one problem; it’s about building a library of successful ways to handle difficult situations based on past failures or unknowns.
Rosa: That ability to adapt based on stored cases, rather than changing the underlying model weights, is a big piece of information here.
Dev: They also summarize the architectural setup: Rasa for dialogue orchestration combined with policy-constrained LLM generation and this strategy memory component.
Taro: And they show that this entire setup supports both text-based and embodied interaction via Furhat to demonstrate grounded support dialogue in a more natural way.
Rosa: Basically, they’re showing how retrieval, safety control, and embodiment can all work together to create a system that actually adapts its communication style over time.
Dev: It’s moving towards systems that are less about predicting the next single word and more about managing a sequence of interactions based on what has worked before.
The paper's improvements: Rosa: The authors suggest improving the system by focusing on making those stored strategies truly reusable and adaptable to new situations, not just directly mapping to the old case.
Dev: They highlight that they are moving beyond simple case retrieval toward a mechanism where alternative responses can be generated and compared against each other dynamically.
Taro: So they are suggesting a system where the agent tries several different approaches based on the stored memories and selects the best one for the current context, which is an improvement over just retrieving a single stored answer.
Rosa: They also emphasize that this entire adaptation happens through retrieval and reuse of these cases and strategies, rather than having to update any of the main model weights.
Dev: That shift from parameter-updating to strategy memory is what makes the online adaptation possible without needing a full retraining cycle for every new scenario.
Taro: If you can adapt behavior this way, it means the system can handle novel student queries in a way that feels more personalized and less like it's following a rigid script.
Rosa: And they also point out that they are exploring how to integrate human-in-the-loop validation for these reusable support strategies in the future work section.
Conclusion: Dev: To wrap up, RPO-PDT is a retrieval-grounded system that uses institution-specific knowledge and policy constraints to provide grounded student support.
Rosa: The key contribution is the reverse-roleplay loop which allows for generating and storing alternative tutor strategies as reusable memory for online adaptation.
Dev: It successfully combines text and embodied interaction with Furhat to demonstrate this adaptive dialogue in a setting that feels more real for the user experience.
Taro: For someone listening, it means we’re seeing how an AI can manage complex support interactions by learning from its mistakes in a safe, constrained environment.
Rosa: It shows how you can build systems that maintain clear role boundaries while still being flexible enough to offer different supportive approaches depending on the student's situation.
Dev: We’re looking at a system that uses stored past interactions to inform future actions, keeping things grounded by those institutional knowledge files and safety policies.
Taro: It’s about creating a support dialogue agent that doesn't just answer questions but learns how to handle the messy parts of student interaction over time.
Filip Janik, Ewa Olton, Robert Smales, Harris Spratt, Shea Tait, Md Zia Ullah, Yanchao Yu
Edinburgh Napier University
cs.RO
Submitted: 2026-06-08
Updated: 2026-10-28
Comments: 3 pages, 2 figures
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 83/100
The gist: RPO-PDT presents a retrieval-grounded, role-play-based dialogue system for adaptive student support in higher education.
Key concepts
- Participant-Aware Role Switching
- This mechanism explicitly tracks whether the current speaker is the tutor or the student. The system uses this information to adjust its conversational approach—for example, changing how it offers advice or questions based on whether it is guiding a tutor or responding to a student.
- Embodied Interaction
- This refers to dialogue that involves physical presence and non-verbal cues, specifically using robots like Furhat. It means the interaction goes beyond just text; the system can use speech, gaze, and facial expressions to make the support feel more human and natural.
- Role Switching State Slot
- The system maintains a specific memory slot that stores the current active role (tutor or student). This is a command-driven mechanism ensuring that the dialogue always knows who it is currently addressing, preventing accidental role confusion during complex conversations.
- Adaptive Dialogue Strategy Generation
- The system generates different support responses by retrieving and reusing previously successful strategies based on the current interaction context. This allows for online adaptation of communicative behavior without needing to fully retrain the underlying model.
Terminology
Summary
RPO-PDT presents a retrieval-grounded, role-play-based dialogue system for adaptive student support in higher education. p. 0 uk
Yanchao Yu
Edinburgh Napier University
Edinburgh, United Kingdom
y. uk
Abstract
We present RPO-PDT: a retrieval-grounded, role-play-based dialogue system for adaptive student support in higher education. uk
Abstract
We present RPO-PDT: a retrieval-grounded, role-play-based dialogue system for adaptive student support in higher education. RPO-PDT is: (1) able to provide institution-specific Personal Development Tutor (PDT) guidance using structured knowledge sources;
(2) constrained by explicit persona, boundary, confidentiality, and
safety policies; and (3) designed around a reverse-roleplay loop
where unresolved interactions are replayed from the student perspective, enabling alternative tutor strategies to be generated and
stored as reusable strategy memory. RPO-PDT is: (1) able to provide institution-specific Personal Development Tutor (PDT) guidance using structured knowledge sources;
(2) constrained by explicit persona, boundary, confidentiality, and
safety policies; and (3) designed around a reverse-roleplay loop
where unresolved interactions are replayed from the student perspective, enabling alternative tutor strategies to be generated and
stored as reusable strategy memory. RPO-PDT supports both textbased and Furhat-based embodied interaction for demonstrating
grounded, safe, and adaptive student-support dialogue.Where the paper says it
Where the paper says it
Where the paper says it
System Architecture and Components
RPO-PDT is implemented using a modular architecture that combines Rasa-based dialogue orchestration, structured institutional knowledge sources, policy-constrained LLM generation, deterministic safety escalation, and an optional Furhat robot interface. The user interaction layer supports two channels: a textbased interface and an embodied spoken interface through Furhat. p. 1 In the following sections, we will
describe each component of our RPO-PDT framework in detail. 1 User Interaction Layer
RPO-PDT supports two interaction channels: a text-based interface
and an embodied spoken interface through Furhat [1]. 2 Dialogue Orchestration with Rasa
The dialogue control layer is implemented using Rasa [3]. Rasa’s
natural language understanding (NLU) module performs intent detection on user utterances, while the dialogue manager tracks conversation state, applies rules and stories, fills slots, and dispatches
custom actions. 3 Rule-Based Role Switching
A key design choice in the current prototype is that role switching is explicit and rule-based rather than inferred automatically. The system maintains an active role state, which is stored as a
dialogue slot, with two possible modes: tutor and student.Where the paper says it
Where the paper says it
Where the paper says it
Knowledge and Policy Management
The system uses structured YAML files as lightweight knowledge and policy stores, including files for module information, course lookup data, university support-service information, PDT guidance, persona constraints, and safety escalation rules. These resources provide factual grounding for queries and encode role and behavioural constraints. p. 1 First, it provides
factual grounding for module, course, assessment, and supportservice queries. Second, it encodes role and behavioural constraints,
such as what a PDT can and cannot do. yaml,’ which classifies user input into
risk levels such as low, high, or crisis. If a crisis-level message is
detected, the system bypasses normal LLM response generation
and returns a predefined escalation response.Where the paper says it
Where the paper says it
Strategy Learning Mechanism
The distinctive component of RPO-PDT is its mechanism for reusing reverse-roleplay strategies. When an interaction is unresolved, the case can be stored in a strategy memory. p. 2 7 Strategy Memory and Reverse-Roleplay
Reuse
The distinctive component of RPO-PDT is its mechanism for reusing
reverse-roleplay strategies. When an interaction is unresolved,
poorly signposted, or identified by the user as needing improvement, the case can be stored in a strategy memory.Where the paper says it
Demonstration Capabilities
The RPO-PDT system is demonstrated through both a text interface and Furhat, allowing users to participate as students seeking support or as tutors exploring alternative responses. Users can test module queries related to academic concerns, wellbeing signposting, role-boundary challenges, and crisis escalation. p. 3 The system routes input through Rasa, retrieves
relevant institutional knowledge, applies persona, boundary, and
safety policies, and generates grounded responses using an LLM. Users can test module queries related to academic concerns, wellbeing signposting, role-boundary challenges, and crisis escalation. RPO-PDT is designed to (1) provide grounded Personal
Development Tutor (PDT) support using institution-specific knowledge, (2) maintain clear role, boundary, confidentiality, and safety
constraints, and (3) reuse effective tutor strategies through a reverseroleplay mechanism in which unresolved cases are replayed from
the student perspective and stored as strategy memory. The current
prototype demonstrates how retrieval, safety control, embodiment,
and role-play-based strategy reuse can be integrated into a practical
conversational interaction system for student support.Where the paper says it
Where the paper says it
Conclusion
RPO-PDT is a retrieval-grounded, role-play-based dialogue system that supports adaptive student-support interaction in a low-resource specialised domain. The system is designed to provide grounded Personal Development Tutor (PDT) support using institution-specific knowledge, maintain clear role, boundary, confidentiality, and safety constraints, and reuse effective tutor strategies through a reverseroleplay mechanism. p. 3 RPO-PDT is designed to (1) provide grounded Personal
Development Tutor (PDT) support using institution-specific knowledge, (2) maintain clear role, boundary, confidentiality, and safety
constraints, and (3) reuse effective tutor strategies through a reverseroleplay mechanism in which unresolved cases are replayed from
the student perspective and stored as strategy memory. The current
prototype demonstrates how retrieval, safety control, embodiment,
and role-play-based strategy reuse can be integrated into a practical
conversational interaction system for student support. The current
prototype demonstrates how retrieval, safety control, embodiment,
and role-play-based strategy reuse can be integrated into a practical
conversational interaction system for student support. Future work
will extend the strategy-memory component, improve automatic
detection of unresolved interactions, and investigate human-in-theloop validation of reusable support strategies.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve adaptive dialogue by implementing a reverse-roleplay loop where
unresolved or weak student-support interactions can be replayed from the student perspective, allowing alternative tutor strategies to be generated, compared, and stored as reusable strategy memory.
This enables online adaptation of communicative behavior without full model retraining. -
Enhance institutional knowledge grounding by utilizing structured YAML files to encode factual grounding for module, course, assessment, and supportservice queries. This allows institutional knowledge and safety policies to remain externally editable without retraining or modifying the underlying LLM.
-
Strengthen safety control by implementing a deterministic layer before response generation that checks both user input and generated responses against keyword-based risk detection. If a crisis-level message is detected, the system bypasses normal LLM response generation and returns a predefined escalation response, ensuring
high-risk situations are not managed solely through generative modelling.
-
Develop explicit role control by maintaining an active role state stored as a dialogue slot with two modes: tutor and student. This command-driven mechanism ensures that the system operates in tutor mode as a Personal Development Tutor (PDT) by default, preventing accidental role drift.
-
Integrate embodied interaction capabilities by supporting both text-based and Furhat-based spoken interaction for demonstrating
grounded, safe, and adaptive student-support dialogue
in a more natural setting. The Furhat bridge connects the dialogue system through an asynchronous integration layer allowing the robot to manage speech input, output, gaze, facial behaviour, and turn-taking cues.
Abstract
We present RPO-PDT, an embodied conversational agent that adapts its conversational role to the current participant. It acts as a Per- sonal Development Tutor (PDT) when supporting students and as a simulated student when faculty rehearse tutoring strategies. Voluntary closed-set participant verification informs a confidence- controlled role manager: stable enrolled-faculty matches activate the Student role, while enrolled students and unmatched partici- pants use the default PDT role and uncertain observations cause no transition. The system combines Furhat interaction, Rasa or- chestration, institutional retrieval, safeguarding, and LLM-based generation. User control is preserved through manual role override, consent-based facial enrolment, temporary interaction identifiers, and explicit retention or deletion of enrolled profiles. The demon- stration exposes both role-switching behaviour and these privacy controls in interaction.
Sources
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation