This project develops algorithms and methods which allow agents to identify and analyze teamwork by the observation of other embodied agents. The goal is to learn from these observations (a) which agents are acting as a team? (b) what collaborative action are they currently executing? and (c) what is the structure of the team, what roles do each of the team members play? The project first creates a corpus of annotated scenarios for training, testing and validation of teamwork learning algorithms. The project then develops a set of algorithms which can detect teamwork between a set of embodied agents by modeling their movement based on reverse force fields. The project also develops algorithms for the robust recognition of known patterns of teamwork using Hidden Markov Models. Finally methods to detect teamwork using the semantic correlation between observations described through semi-formal action descriptions are developed.
The research described in this proposal has immediate practical applications. Recognizing teamwork can help robotic teammates integrate in human teams, with immediate applicability in fields such as disaster response. The analysis of the behavior of the team, as well as successful other teams can be an important feedback in training. In homeland security and surveillance applications, recognizing team action in a crowd can help identify terrorist threats.
