**2.3 Hybrid approach**

In more complex situations, other control architectures appeared (Oreback & Christensen, 2003), including the anticipation in the decision process. This combination is known as hybrid architecture (Arkin, 1990) which has two levels: reactive and deliberative. The deliberative level has to determine and offer the reactive level those behavioural patterns that are necessary for the robot to achieve its objectives. The reactive level has to execute these behavioural patterns by guaranteeing real-time restrictions (Fulgenzi et al., 2008; Fulgenzi, 2009).

Practically, these methods are more complex and require more time calculations. They usually use local methods of obstacle avoidance in order to expand the tree and cover the search space. However, local methods are less suited to dynamic environments (Minguez et al., 2002).

In (Schwartz & Sharir, 1983), a general algorithm was developed. It is doubly exponential time, this limit has been lowered by the Canny algorithm described in (Canny et al., 1990) whose running time is exponential in dimension.

The algorithm D\* proposed by Stentz (stentz, 1995) is a generalisation of the A\* search algorithm in the case of partially known environments. In a finite discretized configuration space, the initial cost of shipping is to move from one configuration to another. A path is first found, and then the robot begins to execute it. If a change in the cost of a path is detected, only the relevant configurations are considered and the optimal path is updated in less time.

Partial motion planning (PMP) (Petti & Fraichard, 2005), the algorithm explicitly takes into account the computation time, constraints and problems safety of navigation in dynamic environment. A tree is grown using a conventional algorithm based on sampling. A node is added to the tree if there is not an inevitable consequence collision. This enables PMP to

To face this limit, deliberative approaches are appeared. They use a global world model provided by user input or sensory information to generate appropriate actions for the mobile robot to reach the target (Pruski & Rohmer, 1997). This kind of approach enables prediction and reasoning in the decision process by considering the current information and the past information (Nilsson, 1980; Sahota, 1994). The deliberative control architecture comprises three modules: sensing, planning and action modules. First, robot sense it's surrounding and creates a world model of static environment by combining sensory information. Then, it employs planning module to search an optimal path toward the goal and generate a plan for robot to follow. Finally, robot executes the desired actions to reach the target. After a successful action, robot stops and updates information to perform the next motion. Then, it repeats the process until it reaches the goal. It can coordinate multiple goals and constraints within a complex environment (Huq et al., 2008;

However, in deliberative navigation, accurate model of environment is needed to plan a globally feasible path. The computational complexity of such systems is generally too great to attain the cycle rates needed for the resulting action to keep pace with the changing environments (it is difficult to obtain a completely known map). To perform necessary

Therefore, these approaches are not proper in the presence of uncertainty in dynamic or real

In more complex situations, other control architectures appeared (Oreback & Christensen, 2003), including the anticipation in the decision process. This combination is known as hybrid architecture (Arkin, 1990) which has two levels: reactive and deliberative. The deliberative level has to determine and offer the reactive level those behavioural patterns that are necessary for the robot to achieve its objectives. The reactive level has to execute these behavioural patterns by guaranteeing real-time restrictions (Fulgenzi et al., 2008;

Practically, these methods are more complex and require more time calculations. They usually use local methods of obstacle avoidance in order to expand the tree and cover the search space. However, local methods are less suited to dynamic environments (Minguez et

In (Schwartz & Sharir, 1983), a general algorithm was developed. It is doubly exponential time, this limit has been lowered by the Canny algorithm described in (Canny et al., 1990)

The algorithm D\* proposed by Stentz (stentz, 1995) is a generalisation of the A\* search algorithm in the case of partially known environments. In a finite discretized configuration space, the initial cost of shipping is to move from one configuration to another. A path is first found, and then the robot begins to execute it. If a change in the cost of a path is detected, only the relevant configurations are considered and the optimal path is updated in

Partial motion planning (PMP) (Petti & Fraichard, 2005), the algorithm explicitly takes into account the computation time, constraints and problems safety of navigation in dynamic environment. A tree is grown using a conventional algorithm based on sampling. A node is added to the tree if there is not an inevitable consequence collision. This enables PMP to

calculations, enormous processing capabilities and memory is needed.

Yang et al., 2006).

**2.3 Hybrid approach** 

Fulgenzi, 2009).

whose running time is exponential in dimension.

al., 2002).

less time.

world.

provide a safe partial path at any time. During the execution of this partial path chosen, another partial path is developed from the end of the previous path developed. This method is applied at real time in dynamic environments, but its major drawback is that it does not consider the uncertainty of the information collected. So it can produce non-executable partial paths.
