A Behaviour Architecture for Cooperating Mobile Robots

The goal of this project is to design an architecture for robot behaviour that can:

• Select low-level reactive behaviour’s for self-preservation
• Plan high-level goal oriented behaviour sequences
• Perform spatial and topological navigation
• Plan cooperative behaviour with other agents

Using a homogeneous selection mechanism that allows sequences of behaviour of any type. The approach is to design an action selection mechanism that incorporates learning. This will be based on an extension of Patie Maes' Spreading Activation Networks. First low-level behaviour is engineered by binding perception-action pairs to achieve reactive robot control. This will give the robots a self-preservation capability in the form of obstacle avoidance. There are a number of different levels of sophistication of cooperation that can be achieved by mobile robots, ranging from simple implicit cooperation - where the robots have no explicit knowledge of each other but their individual behaviour results in a global cooperation; up to high-level cooperation where each robot plans it's cooperative interactions with the other. We are interested in the latter, which necessarily involves communication and also sophisticated purposive navigation by each robot. The cleaning task we have constructed involves one robot ('Flo') scooping litter from near walls and depositing it in piles in open floor spaces. The other robot ('Joh') has a vacuum but cannot vacuum close to walls, hence the necessity for cooperation on the task. Since Joh must vacuum piles of litter left by Flo they must be able to communicate to arrange a location where litter may be placed and collected. The task requirements of the two robots is different and they have different sensors to meet their specific requirements. Below is a picture of Flo, which has 4 custom proportional whisker sensors to meet the need for high-speed close wall following to sweep litter. The sweep (or scoop) can be seen on the right of the picture. Each of the robots also have ultrasonic range sensors and a low-baudrate radio data communications link.



Joh has been fitted with a vacuum to pick up the litter left by Flo. It also has a CCD camerra mounted on top which feeds video signals via a transmitter into a Fujitsu MEP template matching vision system. This allows the development of real-time behaviour for tracking objects, such as Flo, and the recognition of natural visual landmarks, such as doors, litter and general obstacles that should be avoided. Both robots require the ability to select which behaviour to execute next, including which direction to move to achieve robust purposive navigation between locations around the laboratory. To this end we are developing a uniform action selection mechanism that can be used to select low-level behaviours and also high-level behaviours for navigation which incorporates a unified topological and spatial map. The nodes in the map represent locations, which may be associated with natural landmarks via learning. A landmark is a unique feature set at an identifiable location. The features may be in terms of whisker sensor space as it the case with Flo, or visual landmarks as with Joh. Landmarks are joined by topological links that specify which behaviour will move the robot from one landmark to the next. The locations are also associated with a rough spatial position calculated via odometry through learning. The robots communicate via a radio modem data link. The language of communication must be grounded to the sensory environment of the robot in question. This makes communication between robots with heterogeneous sensor spaces difficult. For example, the robots may agree on an identifier for a particular location. When the robots are in close proximity to each other, Joh is capable of seeing Flo using it's vision. In this case the location can be designated. The internal description of the location will be in terms of the sensor space of each robot. So Flo will know the designated location as a relationship to visual landmarks it knows and Flo will know the location in relation to whisker based landmarks. In general a plan for cooperation between two agents can be modeled as a sequence of interleaved behaviours. In summary we wish to:

• Use a behaviour selection mechanism that achieves Integrated Distributed Planning of:
  • High-level goal oriented actions
  • Navigational route planning
  • Cooperation and Communication
• Combine spatial and topological representations
  • with a stochastically learned map that links locations spatially and via behaviour
• Create a physically grounded language for communication and cooperation

The project is currently being investigated by David Jung.

For this project :