FurChat: Artificial Intelligence Applied in Specific Fields

2023-09-14

Large Language Models (LLMs) are advanced deep learning technologies that can interact with humans in real time and answer a wide range of topics. These models have gained widespread attention for their ability to generate human-like answers since the release of OpenAI's ChatGPT.

While LLMs are becoming increasingly popular, most of them are general-purpose and do not provide answers specific to certain topics. On the other hand, chatbots introduced in airports, shopping malls, and public places often rely on other types of Natural Language Processing (NLP) models.

Researchers from Heriot-Watt University and Alana AI have recently created FurChat, a concrete dialogue agent based on domain-specific LLM that can provide information in specific environments. This paper, pre-released on arXiv, introduces how this agent engages in interesting conversations with users through a humanoid robot named Furhat.

Oliver Lemon, one of the researchers, said, "We wanted to study several aspects of embodied AI that can interact naturally with humans. Specifically, we are interested in combining the wide-ranging 'open-domain' conversations that can be had with LLMs like ChatGPT with more useful and specific sources of information, such as information about buildings and organizations (i.e., the UK National Robotarium). We also built a similar system using the ARI robot at the Broca Hospital in Paris for the SPRING project, and it was conducted in French."

The team's primary goal in their recent work is to apply LLMs to context-specific conversations. Additionally, Lemon and his colleagues aim to test the models' ability to generate facial expressions consistent with the communication or response of robots or virtual images at specific times.

Lemon stated, "FurChat combines large language models (LLMs) such as ChatGPT or many other open-source alternatives like LLAMA with a robot that has animated speech capabilities. This is the first system we know of that simultaneously combines LLMs for general conversation and specific information sources (such as documents about organizations) with automatic facial animation for the robot."

The concrete dialogue agent and facial expressions in response are generated by the GPT 3.5 model and conveyed through the Furhat robot verbally and physically.

To evaluate the performance of FurChat, researchers conducted tests with users and asked for feedback after their interaction with the agent. They specifically installed the robot at the UK National Robotarium in Scotland, where it interacted with visitors and provided information about facilities, research affairs, upcoming events, etc.

Lemon explained, "We are researching how to leverage and further develop the latest AI advancements in LLMs to create more useful, usable, and engaging collaborative systems between humans and robots and AI. Such systems require factually accurate information, such as how the presented information is provided in specific documents or images. We are working hard to develop these capabilities to ensure more trustworthy and interpretable AI and robot systems. At the same time, we are developing embodied agent systems that combine vision and language so that they can work alongside humans. This will become increasingly important as human-machine collaborative systems continue to evolve."

In the team's preliminary real-world experiments, the FurChat system appeared to communicate with users in a fluent and informative manner. In the future, this research may lead to the introduction of similar LLM-based embodied AI agents in public spaces, museums, festivals, and other venues.

Lemon added, "We are now expanding the concrete dialogue agent to what is called 'multi-party' conversations, involving interactions with multiple individuals, such as when accompanying relatives on hospital visits. We then plan to extend their use to scenarios where robots and humans collaborate to solve real-world problems."

RECENT AI NEWS

RECENT AI TOOLS