Project : siames
Section: New Results
A model of hierarchical spatial cognitive map and human memory model dedicated to realistic human navigation
Participants : Romain Thomas, Stéphane Donikian.
In the behavioral animation field of research, the simulation of populated virtual cities requires that agents are able to navigate autonomously through their environment. It is of interest to tend to the most realistic human-like planning and navigation. In order to do so, we have designed a navigation system for autonomous agents, which implements theoretical views from the field of human behavior in urban environments. Concerning the perception of the environment, models used in behavioral animation has mainly focused on the visual field to filter what is viewed inside a global geometric database. Information used to navigate has been considered as identical for all autonomous characters and as corresponding to an exact topographic representation of the environment. However, in reality each person has a unique representation of a city map depending on her past experience and on her knowledge of the city. Moreover this cognitive map will evolve with time.
We proposed a new model which allows to represent, for each autonomous agent, an individual hierarchical cognitive map merged with a simple human-like memory model for navigation simulation in an urban environment. It allows to implement navigation as a planned and reactive navigation loop to be computed alternatively.
As shown in Figure20, the system is compounded of 5 different modules:
- The database
which represents the environment and stores all the data related to it. We modeled the urban environment as a database via an a semantically and geometrically Informed Hierarchical Topological Graph (IHT-Graph).
- The cognitive map
which "filters" the information of the environment. Our model of cognitive map has a topological and hierarchical graph structure which partially maps the regions of the environment the agent has explored during the simulation. This map can be seen as a filter on the environment. It does not contain geometrical nor semantic information about the urban objects encountered, but only controls the partial access to the database when the agent recalls or perceives the urban objects.
- The memory controller
which manages the memory in the cognitive map. As a simplified model of human memory, we use the recall and recognition attributes, and their respective thresholds of activation to parameterize the cognitive map in two different ways. The memory model is merged with the cognitive map under two forms: a long-term store mechanism and a short-term one. Concerning the long-term memory, each object of the map is endowed with Recall and Recognition parameters in order to manage the retrieval of information. Links between objects are parameterized through the graph of landmarks which guarantees an associative memory mechanism to the system. The short-term memory respects the Milner rules on its capacity (7±2 elements) and stores subgraphs of the cognitive map, linked together in a mereo-topological way.
- The route planning module
which implements the planning and navigation algorithms. The agent plans its route using its cognitive maps, knowing that most of the time, the navigation plan is not complete enough to reach its destination, it has to switch between two different types of navigation: planned navigation and reactive navigation.
- The navigation module
based on the HPTS decisional system which manages the behaviors of the agents in the environment, which underlies the composition of the planned and reactive navigation algorithms. The first one uses the plan generated using the information gathered in the cognitive map, while the second one, holds for the navigation when the agent is lost (i.e. the agents fell into the critical decision cases ensuring a local loss in the environment). The reactive navigation mode, following first main axes, allows the agent to plan a new route taking into account the new information gathered. During the navigation, the pedestrian meets relevant urban objects not recalled but recognized, which trigger the recognition of a region of the map.
