NarraGuide: LLM-Powered Narrative Telepresence
- NarraGuide is an LLM-based narrative mobile robot that fuses telepresence, SLAM-driven autonomous navigation, and location-aware dialogue to guide users through remote environments.
- It integrates a robust system architecture—TurtleBot3, ROS2, LiDAR-based SLAM, and GPT-4—to support both navigational control and context-rich conversational interactions.
- Field deployment in a geology museum demonstrated enhanced user trust and learning, with participants significantly preferring the robot’s autonomous suggestions despite control and grounding challenges.
NarraGuide is an LLM-based narrative mobile robot for remote place exploration that combines telepresence, autonomous navigation enabled by SLAM, and location-aware conversational capability powered by an LLM so that a remote user can explore and learn about a real place through dialogue rather than through camera control alone (Hu et al., 2 Aug 2025). Developed as a prototype and deployed in a geology museum, it links a mobile robot’s physical location to exhibit-specific narrative material, allowing users to ask questions, receive contextual introductions, and alternate between guided touring and direct verbal control (Hu et al., 2 Aug 2025).
1. Conceptual scope and motivation
NarraGuide addresses a specific weakness of conventional robotic telepresence in unfamiliar environments. Prior remote-exploration technologies can let people access places they cannot physically visit, but they often leave the user to pilot the robot, decide where to go, determine what is interesting, and infer what the camera is showing. In unfamiliar settings, effective navigation and situational awareness depend on prior knowledge of the environment, including layout, points of interest, and the meaning of nearby objects or exhibits. NarraGuide is proposed as a response to that burden: the robot should know where it is, retrieve information relevant to that location, and converse with the user using natural language so that exploration does not require prior familiarity with the site (Hu et al., 2 Aug 2025).
The system is framed not as a teleoperated camera on wheels and not as a standalone chatbot, but as a narrative mobile robot that unifies telepresence, location awareness, dialogue, and narrative guidance. The paper illustrates the intended interaction with a park scenario in which a remote user asks about a favorite bench under oak trees, the robot suggests a nearby oak tree, navigates there, and narrates facts such as oak lifespans. In the museum deployment, the same design logic becomes exhibit-specific: the robot can proactively introduce nearby exhibits, answer questions, suggest where to go next, and narrativize what the user is seeing in real time (Hu et al., 2 Aug 2025).
2. Embodied system architecture
NarraGuide is organized into three main modules: User Interface, Dialogue System, and Robotic System. The architecture separates remote interaction, language processing, and embodied autonomy while keeping them tightly coupled through the robot’s current location and the user’s spoken input (Hu et al., 2 Aug 2025).
| Module | Function | Main components |
|---|---|---|
| User Interface | Dialogue-based remote interaction | Chrome extension, speech-to-text, Chrome TTS, ROS bridge |
| Dialogue System | Intent classification and response generation | GPT-4 prompts, intent handlers, function calling, context retrieval |
| Robotic System | Mobile telepresence and navigation | TurtleBot3 Waffle, LDS-01 LiDAR, IMU, wheel odometry, ROS2, Nav2 |
The robotic platform uses a TurtleBot3 Waffle equipped with a LiDAR sensor (LDS-01), an IMU, and wheel odometry. A smartphone is mounted on the base using an adjustable rod and a gimbal stabilizer, and the phone hosts a Zoom teleconference call that provides the live audiovisual feed to the remote user. ROS2 integrates sensing, navigation, and communication. On the user side, the interface is implemented as a Chrome extension overlaying a teleconferencing page; user speech is transcribed by a speech-to-text module, robot responses are rendered through a Chrome TTS module, and ROS bridge carries data between browser and robot (Hu et al., 2 Aug 2025).
3. Semantic mapping, exhibit grounding, and dialogue-guided navigation
A crucial part of NarraGuide is semantic place awareness rather than localization alone. The museum map is built with LiDAR-based SLAM using Cartographer, with LiDAR, wheel odometry, and IMU fused to improve mapping accuracy. After map creation, exhibit locations and exhibit numbers from the original museum map are aligned to the robot map, an optimal viewing pose is manually determined for each exhibit by two authors, the map is segmented into grids, grids are clustered into general areas, and nearby exhibits are grouped according to the grids they occupy. This lets the system infer the current grid, current general area, and nearby exhibits from the robot’s pose (Hu et al., 2 Aug 2025).
The narrative layer is built from museum-specific data. Two members of the research team labeled material from two museum tour guidebooks and exhibit labels by exhibit number and categorized it into history, activities, and miscellaneous information. The resulting dataset covered all 68 exhibits in the museum. GPT-4 then generated two additional fields for each exhibit: an introduction for each exhibit and sample dialogues between a museum visitor and a tour guide. These exhibit introductions, nearby-exhibit descriptions, and area descriptions are inserted into intent-specific prompts at runtime (Hu et al., 2 Aug 2025).
The dialogue system first classifies user input into conversational intents or navigational intents. The conversational intents are inquiry about the museum, comment, and free chat. The navigational intents are low-level robot control and high-level robot control. Low-level commands are mapped to turn left, turn right, move forward, and move backward. High-level commands are processed with GPT-4 function calling to extract a goal such as an exhibit number or a next-stop request derived from visit history and predefined exhibit order. For conversational turns, the prompt includes the introduction of the current exhibit, sample dialogues for the current exhibit, the current area, and introductions of nearby exhibits. If there is 45 seconds of silence, the robot may proactively initiate social dialogue (Hu et al., 2 Aug 2025).
The system’s interaction style is explicitly exhibit-grounded. In one example, the robot begins with: “Let’s start with our first exhibit, which showcases a beautiful piece of red granite. Can you tell me what type of rock this red granite is?” and then elaborates that the specimen originated from the vicinity of Wausau, Wisconsin, and is estimated to be approximately 1.75 billion years old. In another example, when asked “Tell me more about Galena,” the robot answers that Galena is formed in a wide range of hydrothermal environments and is a component of some granites. These examples show that narrative guidance is implemented as contextual explanation tied to a current exhibit rather than as free-form small talk (Hu et al., 2 Aug 2025).
4. Museum deployment and quantitative findings
The prototype was deployed in a local geology museum. Although the full museum contained 68 exhibits across four areas, the study focused on about half the museum, comprising 23 exhibits across three zones, with 11 exhibits selected in consultation with an experienced geologist based on local relevance and popularity. The user study involved 20 participants, with ages 19–77, mean age , and standard deviation . Each participant used NarraGuide to explore the museum remotely for 20 minutes and was asked to visit at least five exhibits. After a tutorial and practice with both conversational interaction and directional commands, participants completed the tour, a semi-structured interview, and a questionnaire (Hu et al., 2 Aug 2025).
The quantitative analysis reported three paired-sample t-tests.
| Comparison | Means | Result |
|---|---|---|
| Acceptance vs rejection of robot suggestions | acceptance , ; rejection , | |
| High-level vs low-level commands | high-level , ; low-level , 0 | 1 |
| Museum inquiry vs robot control speeches | inquiry 2, 3; control 4, 5 | 6 |
The only significant difference was that participants accepted the robot’s suggestions more often than they rejected them. By contrast, there was no significant difference between high-level and low-level robot control, and no significant difference between museum inquiries and robot control commands. The result indicates that users treated NarraGuide as both a guide and a controllable robot, rather than favoring one interaction mode exclusively (Hu et al., 2 Aug 2025).
5. Perceived roles, learning effects, and public-space interaction
Participants did not experience NarraGuide in a single role. They variously described it as a guide, a tool and medium, a companion, and a representation of self. Many participants saw it as a knowledgeable, intelligent, professional guide, especially when they were unfamiliar with geology or with the museum. Some preferred asking the robot questions over asking a human guide because they felt less judged. Others framed it as a learning tool or “Wikipedia,” and some described it as a “proxy” or as their “eyes” in the museum. A smaller number emphasized companionship and loyalty, while several moved between guide-like and embodied interpretations during the same session (Hu et al., 2 Aug 2025).
The dialogue was often experienced as pedagogically productive. Participants reported that the robot’s questions made them think more about the exhibition, slow down, become curious, and recall forgotten geology knowledge. Many felt that the live question-and-answer format helped them learn things they would not have known otherwise, and some preferred it to searching on a phone or listening to a pre-recorded guide. At the same time, the study exposed control and grounding problems. Some users appreciated the robot’s initiative because it reduced uncertainty, but others wanted confirmation before autonomous moves, more alternatives, and better reminders of what had already been visited. Voice commands were awkward for precise movement, the robot could not listen while speaking, and speech recognition could suffer from bystander voices and museum noise (Hu et al., 2 Aug 2025).
The strongest content-related limitation was referential grounding. When a display case contained multiple specimens, users often could not tell which object the robot was describing, and they requested pointing gestures, camera centering, zooming, visual labels, or supplementary images. Trust complicated this further. Most participants reported that they trusted the robot’s explanations, yet the authors observed hallucinations, including cases in which the robot fabricated answers and participants still believed that it had correctly identified a colored rock. The paper therefore identifies a specific risk: fluent narrative explanation can produce unwarranted trust in inaccurate content (Hu et al., 2 Aug 2025).
The deployment also highlighted social effects of operating in a public museum. Interaction logs and recordings were analyzed alongside YOLOv8 human detection and manual annotation of bystander encounters. Some children danced in front of the robot, hid from it, followed it, or blocked it. Some participants enjoyed the resulting sense of social presence, while others found bystanders distracting because of view blockage, added noise, or concern about being overheard. Preferences about self-presentation were likewise mixed: some participants wanted their face displayed on the robot, while others preferred privacy or suggested avatars or emoji-like representations (Hu et al., 2 Aug 2025).
6. Limitations and placement within a broader research trajectory
The paper identifies several technical and design limitations. The current prototype uses a 2D semantic map, and the authors explicitly suggest that future systems could use 3D spatial information so that the robot can reason about relations such as “the fossil on the right.” They also point to the need for stronger trustworthiness under hallucination risk, more explicit confirmation before autonomous moves, more options rather than unilateral suggestions, support for customized commands, stronger visual interfaces such as zooming and highlighting, possible mixed-reality overlays, and safer LLM-based planning methods for robot behavior (Hu et al., 2 Aug 2025).
A plausible implication is that NarraGuide belongs to a broader class of systems in which narrative guidance supplements navigation or retrieval rather than serving as ornament. StoryNavi, for example, constructs “customised video play experiences by retrieving materials from original videos” and organizes those clips into a cohesive narrative rather than returning timestamps alone (Xu et al., 2024). DialNav likewise formalizes a collaborative navigation setting in which a remote Guide must infer the Navigator’s location through multi-turn dialog, making communication essential for task success rather than optional (Han et al., 16 Sep 2025). This suggests that NarraGuide’s distinctive contribution lies in bringing that narrative-guidance logic into embodied telepresence: the guide is not merely explaining a place, but actively coupling explanation, spatial context, and locomotion in real time (Hu et al., 2 Aug 2025).
