Simulation and Coordination of Mobile Robots
Consider the scenario in which we placed a 3D environment scanner into an arbitrary room with various everyday objects (e.g., furniture) and produced a 3D model of what is in the room.   Now letting thousands of small cube-shaped robots into an empty room which has the same proportions as the scanned room.   We are investigating algorithms in which the robots will autonomously re-build a physical 3D representation of the model that was scanned in.   This is similar to the idea of automatically having the robots coordinate and merge together to build a physical lego-like (or minecraft-like) structural representation of the actual room with the furniture.   The robots need to be designed in a manner in which they can flop around as well climb one another and attach themselves together in a manner that is structurally sound.   We are working on simulations of such robots at this time.



Mobile robots can be used to solve various real-world problems without human supervision.   Teams of robots are sometimes required because the a single robot may not have enough power or agility to manipulate large heavy objects such as pushing heavy items in a factory, transporting dangerous goods, or pushing a boat into a port.   We have been researching algorithms to coordinate a team of autonomous mobile robots to solve the task of pushing an arbitrary 2D polygonal-shaped fixed-weight object from one location to another in a 2D environment with various polygonal objects.   The problem requires the robots to place themselves around the push-object efficiently so as to be able to transport the object smoothly and as quick as possible along a computed path trajectory.   Depending on the number and placement of the obstacles, the problem usually requires the push-object to be rotated at various times during the transportation (much like having to adjust the orientation of a couch being moved into an apartment).   The robots must have a strategy for re-positioning themselves efficiently around the push-object as well as having a strategy to coordinate the efforts of all the robots so that the push-object is always moving in the correct direction.


When teams of robots are assigned to lift a large heavy object (e.g., for transportation purposes), it is important that the weight is distributed among the robots.   Proper weight distribution will help to minimize the number of robots required to lift the object while ensuring that the robots are less likely to be overburdened, which could lead to hardware failure as well as increase the likelihood of dropping the object being transported.   We have been investigating algorithms for distributing the weight of a 2D polygonal uniformly-weighted object by determining efficient placement of a team of robots based on the center of gravity of the object.



Individual mobile robots are limited in their abilities to accomplish tasks that require coverage of large areas.   For example, if a single robot had to search for lost person/device in a terrain, it would be quite difficult and time consuming.   However, if a large team of robots was used for the search task, lost person/device would be found much more quickly, provided that the robots are coordinated in their search efforts.   We performed research into ways of organizing simple robots (i.e., very few sensors) into hierarchical arrangements whereby small teams of robots are managed by a team coordinator at the lowest level.   These coordinators are themselves managed by a higher-level coordinator, and so on.   At the top level, a human can be coordinating the highest level of coordinating robots.   This allows global knowledge to trickle down through the hierarchy so that the individual robots can be directed to various areas of the environment more efficiently and timely.  Communication is simplified because it only occurs between the individual robot and their coordinator, who directs their efforts.   Therefore, if any communication failure occurs within the hierarchy, the lower level coordinators can still function, given their latest search directive.


In this project, we developed an algorithm to allow a robot with simple sensors (e.g., whiskers and a compass) to create a map of an unknown 2D environment with obstacles.   The robot is able to trace out the border of the environment as well as each individual obstacle to produce polygonal representations of the environment.   Since the information comes only when travelling along obstacle borders, there is no information available while the robot is travelling in the open spaces of the environment.   Hence, "theoretical bridge" connections are introduced between various polygons to obtain relative positioning between objects.


This project simulated a colony of robots that were able to interact together to perform various tasks.  One task that was implemented was that of coordinated mapping.   Each robot produced a portion of the environmental map and then these maps were merged together to produce a global map.  


A 4th year robotics course here at Carleton made use of a limited number of PropBots (i.e., robots designed for the course) which were used within our robotics lab test environment.   In order to test robotic algorithms outside of the lab setting, this software simulator was implemented which allowed students to work at home to test their software in simulated robots before coming in to the lab.   The simulated robots were developed to closely correspond to the actual PropBot robots in regards to movements and sensors.


This simulation tool was developed to simulate colonies of robots that use very simple instinctive neuron networks as those in my masters thesis robot.  This tool allows a user to create different kinds of cleaning robots that can clean near well lit areas, near other robots and near both.  I created colonies with different behaviours by varying the types of robots such as swarm leaders, followers and light directed cleaners.  By observing the behaviour of these robots I was able to determine which types of robots performed well together.  A more important result is that the robots are able to perform well together even with minimal sensor complexity.



The robots have very simple sensors such as detecting light and signals from other robots.  It allows the user to investigate different kinds of swarm behaviour and see how they apply to simple tasks such as cleaning an environment.  The program has nice interface that allows the user to add sensors and actuators to the robots and code their behaviour at run-time.  In addition, users can design instincts as networks of neurons by using a graphical drag and drop interface.  The system allows the user to make libraries of robots, sensors, actuators and instincts.  The screen snapshots below show the dialog boxes for creating robots, sensors and instincts:






This project investigated the various mechanisms that motivate behaviours in A.L. (artificial life).  The system simulates a colony of bugs whose main goal is to remain alive.  The bugs have attract/repel behaviours to various kinds of stimuli (such as different foods, bugs of the same or different sex, light sources and gravel patches).  The bugs are able to do multi-level associative learning in their behaviour hierarchy which allows them to stay alive.



The user can add multiple bugs to the environment as well as various types of stimuli. The user can poison certain foods and or zap bugs when they are doing something considered "bad". For example, the user can stop male bugs from chasing female bugs or stop a bug from eating apples that are on gravel patches near light sources. Some very interesting behaviours emerge from varying the stimuli and punish patterns. As a bonus feature, there is a brain browser (see image below) that allows the user to examine the brain of a bug to view the hierarchical behaviour mechanisms that led it to behave in a certain manner as well as observe the associative learning that has taken place.


This project involved implementing an insect that learns how to coordinate its legs to achieve walking by using back propagation neural networks.  Each leg has its own network that indicates whether the leg should be in stance mode (down and pushing backward) or swing mode (up and swinging forward).  The network of each leg connects to the adjacent legs across and behind it.  When the insect falls down (i.e. not enough legs are in stance mode or the center of gravity is off balance), then the leg networks are "punished".  The leg networks are all rewarded when the insect makes forward motion.  When created, the insects are endowed with a certain amount of pre-learnt instincts.  After about 20 seconds (which varies depending on the amount of instinctive learning), the insect is able to coordinate its legs and walk.  The user can then pluck off some of the legs and watch as it attempts to compensate for the lost limbs.  As an additional feature, the insect has two back propagation networks linked to light sensors.  The insect is able to quickly learn how to seek out light sources.