This is a plan recognition system targeted at automotive traffic. It is generally capable of handling continuous domains where multiple agents interact with each other and where plan recognition needs to carried out on-line.
If you are primarily interested in the Golog interpreter, click here.
Our approach to plan recognition is as follows. (See this paper and these slides for details.)
- First the domain is modeled in terms of a situation calculus basic action theory which basically defines the available primitive actions and fluents. (This paper and these slides present our theory for reasoning about traffic situations.)
- Then one defines a plan library which consists of Golog programs. Each of these programs represents a typical behavior pattern.
- To perform a concrete plan recognition task, execution of these programs is simulated. The effects in this simulation are compared to the observations of the real world. If the observations and the simulation are consistent, the respective program is considered a potential explanation of the agent's action, i.e., a recognized plan.
Our current application domain is automotive traffic. We use TORCS as driving simulation. It transmits observations to the plan recognition twice a second. The modifications to TORCS can be found in this repository. For testing purposes, observations can be recorded and then replayed (using scripts/replay.c).
Contact: Christoph Schwering (schwering at gmail dot com).