HRI Furhat — a social robot for student mental-health check-ins

A Furhat social robot that runs a short, empathetic, LLM-driven mental-health check-in with university students. It greets the student, asks a sequence of check-in questions, listens to spoken answers, reads the sentiment of each response, and replies with brief empathetic acknowledgements — with facial gestures and expressive LED cues throughout.

Built for a Human-Robot Interaction course project at Uppsala University.

What it does

• Spoken dialogue on a Furhat robot — connects to the Furhat real-time API, configures the voice, and drives a greet → ask → listen → respond loop with LED state cues (main.py, furhat_client.py, furhat_controller.py).
• LLM-driven empathetic responses — a system prompt frames the robot as a social agent conducting a 3-minute student check-in, and each answer gets a short, empathetic acknowledgement (LLMmodule.py, prompts.py).
• Swappable LLM backends — Groq and Google Gemini sit behind a common interface in LLMmodule.py, and main.py picks one via the LLM_BACKEND env var (default groq). A separate OpenAI path generates the empathetic/neutral replies in furhat_controller.py.
• Sentiment sensing — each student response is scored with a RoBERTa sentiment model (cardiffnlp/twitter-roberta-base-sentiment-latest) so the interaction can respond to how the student is feeling (SentimentEmoModule.py).
• Facial gestures — expressive gestures are triggered during the interaction (facial_gestures.py).

Architecture

File	Role
main.py	Orchestrates the greet → ask → listen → respond loop
furhat_client.py / furhat_controller.py	Furhat real-time API connection, voice, listen/speak, LED cues
LLMmodule.py	LLM dialogue (Groq / Gemini) and the check-in logic
openai_helper.py	OpenAI backend
SentimentEmoModule.py	RoBERTa sentiment scoring of student responses
facial_gestures.py	Furhat facial-gesture triggers
prompts.py	System and turn-level prompts for the check-in

The study it was built for

The system was the apparatus for a between-subjects HRI study: does empathetic robot behaviour change how a social robot is perceived during a wellbeing check-in? Two versions of the check-in were produced from identical scripts — an empathetic one (supportive language, smiles, eyebrow movement, reflective nods, attentive gaze, slowed prosody) and a neutral one (task-focused acknowledgements, steady gaze, no emotional expression).

Twenty-three Uppsala students were randomly assigned to one condition (empathetic n = 12, neutral n = 11) and rated the robot on the Robotic Perceived Empathy (RoPE) scale and the Godspeed likability/intelligence subscales. Mann–Whitney U tests found no significant scale-level difference between the conditions (RoPE U = 70.0, p = .49; likability U = 69.5, p = .34) — a useful reminder that for a short, clearly-structured check-in, a simpler neutral style can be about as acceptable as an expressive one.

Team and my contribution

Group project for 1MD043 Human–Robot Interaction (Uppsala) by Dan Zhang, Venkatesh Akhouri, Rafael Proença, and Sami Mnif.

My part: I assisted with the experimental design and the report, helped structure the interaction flow and the analysis, and added a Google Gemini backend to the system as an alternative to the OpenAI/Groq path.

Running it

You need a Furhat robot (or the Furhat SDK desktop launcher) and API keys in a .env file at the project root:

FURHAT_AUTH_KEY=<your-furhat-key>
GROQ_API_KEY=<your-groq-key>        # for the Groq backend
OPENAI_API_KEY=<your-openai-key>    # for the OpenAI backend
GEMINI_API_KEY=<your-google-ai-key> # for the Gemini backend
LLM_BACKEND=groq                    # main.py dialogue backend: groq (default) or gemini

pip install -r requirements.txt
python main.py            # set LLM_BACKEND=gemini to run the Gemini backend instead

Keys are read from the environment — nothing is hard-coded. This repo is a fork of the group's shared project.