Context-aware Intent Prediction with Self-Healing for Improving Human Machine Cooperation in Support of Complex Collaborative Decision Making
Abstract
The ability of human beings to accurately recognise others’ intents is a significant mental activity that involves reasoning about intents, such as, what other people are doing, why they are doing it, and what they will do next. Therefore, the quality of interpersonal and human-machine communications can be enhanced through employing such intent-based intent reasoning to identify other human beings as peers, competitors, or bystanders. The above observation can be utilized to boost the capability of human-machine interactions which serve as the core of interactive intelligent systems (e.g., robot, logistics units or other sophisticated military systems). Such systems aim to meet the increasingly complicated defence demands, namely, improved operators’ performance and training techniques, autonomous weapons, reduced casualty, post-casualty recovery, and mental health management. Previous research has revealed that human intents could be inferred by measuring human multi-faceted activities from multiple heterogeneous information sources, such as, body and brain sensors (e.g., sensors for detecting physiological signals like ECG, brain signals like EEG and gestures/eye movement/voice from IMU sensors like accelerometers and gyros etc.), along with implicit and explicit contexts (e.g., presence, environment and activities). Most previous research focuses on detecting the intent other than prediction, or recognize the intents which have been predefined, where only partial observations or few clues of the predicted intent have been observed or newly unseen intent recognition, have not yet been explored [5-11][16][18][29]. Future intent prediction is crucial in real-life scenarios, where anticipatory response is required such as active sensing and autonomous navigation to make responses actively. For example, it can help autonomous vehicles to decide how to manoeuvre depending on the next predicted intent or assist robots to make future decisions. In these scenarios, existing systems can only detect the intent when it has already occurred or partially occurred, which cannot give operator sufficient time to respond. The aim of this project is to develop theoretical foundations and a data-efficient intent prediction paradigm that can (1) capture human-machine interactions and infer implicit contexts; (2) analyse and predict future intents from both explicit and implicit contexts; (3) discover and recognize newly unseen intents with limited examples; and (4) provide effective mitigation strategies to improve performance, such as adjusting levels of automation, or adapting visualizations.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 23, 2019
- Source ID
- N629091912009
Entities
People
- Lina Yao
Organizations
- Office of Naval Research
- United States Navy
- University of New South Wales