hwu-social-alignment-robots-2025 / alignment-social-robots

This project investigates human-robot alignment using a Furhat robot. Participants rank items, discuss with the robot, and re-rank to explore alignment dynamics. The study assesses whether alignment depends on concept nature and its correlation with trust in robots, contributing to human-robot interaction insights.

Alignment Social Robots

Overview

Our project explores how humans and an anthropomorphic​ robot, powered by a sophisticated Large Language Model​ (LLM), achieve alignment in conversations.

To measure the alignment, we will be running experiments of human-robot interactions with a Furhat robot. Asking the participants to complete a ranking tasks before and afterward to quantify the robot influence over the subject.

You'll find here the implementation of the Furhat client that will make the connection between the Furhat robot and the LLM API.

Quickstart

To run the Furhat client please follow the instructions of the quickstart documentation.

Our Team

This project was made by:

• Aswin Shaji
• Guilhem Santé
• Jake Peterson
• Jeff Sherer
• Marc Puig Arocas
• Maximilien Bildstein
• Mizan Haque
• Prudhvi Venkata Sai Paleti
• Robbie McPherson
• Sattwik Mohanty

Under the supervision of:

• Alessandro Sugglia
• Marta Romeo
• Greta Gandolfi

alignment-social-robots's Issues

Robot stop listenning if participant talk for too long

Currently, the robot have some issues during its listenning, It stop listenning at some point if the particpant talk for too long.

After reading a bit the documentation, I’ve found it is actually a parameter we could actually change:
https://docs.furhat.io/listening/

Add continous export dialog history

Using ChatGPT, resume the dialog history to have context of the previous dialog history. In this way we can manage the conext window minimazing the losing information.

Structure the prompt engineering parameters

As we'll be trying a lot of prompt engineering options, we should make the switch from the one prompt to another as simple as possible. Beginning by structure the prompt engineering parameters, an potentially add the loading of configuration file for better integration.

Add a procedural generation of the robot ranking

In order to keep some randomness during each experiments the robot should be capable of generating its own unique ranking for each participants.

Change queryied OpenAI model

We’re currently using querying the gpt-3.5-turbo-instruct as the LLM to generate our responses. But this model only have a context window of 4,096 tokens, which is no a lot compared to some other model like gpt-4-turbo-preview which can take over 128,000 tokens (and also the training data is more recent).