CogHRI
Cognitive Communication as Moving Target Tracking
CogHRI developed self-improving, bidirectional human-robot interaction technology capable of understanding users’ complex emotional states and transactional intent through continuous interactions — targeting a 95% successful response rate.
Overview
The core objective of CogHRI was to develop autonomous, bidirectional HRI technology that progressively understands and learns a user’s combined cognitive and emotional states, as well as their transactional intent (current intent, future sequential intent, and intent transitions) through sustained robot-user interaction. The system aimed to respond appropriately in 95% or more of such interactions.
Cognitive communication is framed as a moving target tracking problem: the user’s intent and emotional state are constantly shifting, and the robot must continuously update its model through a perception-action-learning cycle with dynamic Bayesian inference.
The project addressed this challenge through five core HRI element technologies, collectively named MESSI:
| Sub-task | Domain |
|---|---|
| HRI.Map | Robot spatial mapping and navigation |
| HRI.Emotion | User emotion recognition |
| HRI.Service | Service planning and execution |
| HRI.System | Full system integration |
| HRI.Intention | Dialogue (linguistic) intent understanding |
SNU Biointelligence Lab led the HRI.Intention sub-task (Sub-task 3), focusing on dialogue understanding and generation based on a perception-action-learning cycle, and developing real-time conversation models backed by hierarchical probabilistic inference to verify and support transactional intent recognition.
Research Team
Principal Investigator: Prof. Byoung-Tak Zhang (HRI.Intention sub-task)
Researchers:
- Ha-Young Jang
- Beom-Jin Lee
- Jin-Hwa Kim
- Chung-Yeon Lee
- Kyung-Min Kim
Technical Approach
The HRI.Intention sub-task pursued four lines of technical development:
- Dialogue corpus collection and machine-learning-based dialogue modeling — building annotated corpora of human-robot dialogue and training statistical models of conversational structure.
- Real-time dialogue understanding with dynamic model refinement — applying learned dialogue models to live interactions and updating the model incrementally as new exchanges occur.
- Linguistic intent recognition using nonverbal transactional cues and complex emotion — integrating body language and affective signals to improve interpretation of spoken intent.
- Support module for nonverbal transactional intent verification — providing auxiliary inference to cross-validate cues from non-linguistic channels.
The overall goal was to maximize transactional-intent accuracy while improving dialogue coherence, comprehension depth, and processing speed through real-time dialogue understanding and generation models.
Collaboration
| Partner | Role |
|---|---|
| Hanyang University (Prof. Il-Hong Suh) | HRI.Map, HRI.Service — overall project PI |
| Kyungpook National University (Prof. Min-Ho Lee) | HRI.Intention support |
| KAIST (Prof. Dong-Soo Kwon) | HRI.Emotion |
| KIST (Dr. Seung-Jong Kim) | HRI.System integration |
Publications
- J.-H. Oh, H.-S. Chun, and B.-T. Zhang, “Generating cafeteria conversations with a hypernetwork dialogue model,” in Proceedings of the 14th International Symposium on Advanced Intelligent Systems (ISIS 2013), pp. 1424–1435, 2013.
Contact
Contact: Ha-Young Jang Email: hyjang at bi.snu.ac.kr