Introduction
Negotiation research has traditionally been bifurcated into two distinct domains: automated negotiation (agent-to-agent) and human-agent negotiation (HAN). While platforms like GENIUS and NegMAS have standardized agent-to-agent research, HAN researchers often face a significant "startup tax." Developing a system for a human to negotiate with a robot requires the simultaneous integration of speech-to-text, natural language parsing, emotional state tracking, and physical robotic control. NEGOTIATOR addresses this challenge by providing a modular, open-source framework that handles these complex interaction layers, allowing researchers to focus on core negotiation strategies and psychological outcomes.
System Architecture
The NEGOTIATOR framework is built on a modular architecture centered around a "NegoCore" orchestrator. Utilizing Docker containers and RabbitMQ for message passing, the system ensures that individual components—such as perception, emotion modeling, or agent logic—can be swapped or updated independently without rewriting the entire software stack.
Key modules include:
- Perception: Handles Speech-to-Text and Text-to-Offer parsing to convert natural language into structured bids.
- Emotional Modeling: Recognizes human emotions from facial expressions or text, supporting categorical, dimensional, and continual models.
- The Agent: Comprises a three-part "brain" consisting of an Offering Strategy, an Acceptance Strategy, and an optional Opponent Model.
- Interaction Manager: Translates agent decisions into human-friendly outputs, including robotic gestures (pre-configured for Nao, Pepper, and QT robots), argument generation, and Text-to-Speech.
User Interface and Interaction
The framework includes a web-based Graphical User Interface (GUI) designed for both chat-based negotiation and experimental monitoring. This interface provides real-time feedback on preferences, utility, and emotional states, facilitating a more transparent interaction for the human participant.
A unique feature of NEGOTIATOR is its multimodal "mood" transfer. The agent does not merely send a bid; it transmits a "Mood" alongside the offer. This allows for the synchronization of verbal strategies with non-verbal cues, such as a robot appearing frustrated when rejecting a low-value offer or excited when reaching a compromise.
Applications and Results
The versatility of NEGOTIATOR has been demonstrated across several successful use cases:
- Gesture Impact Studies: Investigating how a robot's physical gestures influence human negotiation behavior.
- Emotion-Aware Bidding: Using live video processing to adjust agent strategies based on the human's detected emotional state.
- Embodiment Comparisons: Comparing human reactions to physical robots versus virtual avatars in identical negotiation scenarios.
- Cross-Platform Adaptability: Rapidly deploying the system across different hardware (Nao, QT, Pepper) and software interfaces.
Conclusion and Future Work
NEGOTIATOR significantly lowers the barrier to entry for Human-Robot Interaction (HRI) and HAN research by providing a standardized "plumbing" for complex interactions. While currently focused on bilateral negotiations, future development aims to support multilateral scenarios and deeper social awareness. This framework paves the way for more sophisticated studies into the psychological and social dimensions of human decision-making in negotiation.
Summary updated