SketchaBot is a one of a kind drawing and painting robot. He uses light sensors to detect the paint and dip his brush. On the back of SketchaBot is a complex holster that holds the writing utensil, whether it be a paintbrush or a marker. The holster uses one motor and two touch sensors to sense when the robots utensil has reached the writing surface or when it has been retreived all the way. Precise timing is key with the programming if this robot. By changing the timing intervals just a tiny bit in the beginning, it would throw off the whole picture.
Marcel was designed to use no programming (besides .On FWD.), and no sensors to detect obstacles and take an action to avoid them. I used a series of gears including a differential gear to switch the direction of the right wheel when an obstacle obstructed the robot.s path, making it turn right to avoid the obstacle. The robot has a long rail across the front that gets pressed down when it bumps into an object. This rail is connected to a mechanism that lifts two gears off of the differential and the right wheel axle, and engages a stationary gear below the differential to force the gears inside the differential to turn, reversing the right wheel; this process happens as long as the front rail is depressed. When the rail releases, the gears return to place, the stationary gear moves away from the differential, and the robot resumes going forward.
Two robots learn to following a complex black line using genetic algorithms. They also shire information and knowledge as learning. Co-operative learning is an advantage of this project. Through the limited RI communication, these two robots, Dee and Dum, have successfully learned following the black line. Dee and Dum use three individuals, meaningful crossover operator, and frequent mutation. Testing results show 7 out of 8 trails are successful.
This sumo-bot, Rocky, is not the heaviest, biggest, or fastest, rather he relies more on other abilities given he didn't have as many motors as many of the other sumo-bots. Rocky relies on a sturdy construction to absorb blows and pushing from other sumo-bots without breaking. He also features counter rotating big toothed front wheels that spin just off the track. These wheels are used to grab and flip over any opponents. Rocky also is programmed to perform evasive maneuvers if he detects he is about to be pushed off the track. Once the evasive maneuvers have been performed, he attempts to immediately push or flip his opponent off the track, while the opponent is still close to the edge.
The QWalking Lego robot employs the Q-learning technique to learn how to walk with two independently controlled motors. It uses two rotation sensors. The robot collects the current state of the environment by looking at how far both rotation sensors have gone. Then the robot looks at the Q-table and find the most rewarding action. If this action keeps the robot walking synchronized, a reward is given to the robot. Otherwise, a punishment is given to the robot. Either way the Q-table is updated. This technique forces the robot to keep the both legs synchronized.
This Lego robot is a simulation of fish counting device. It uses an artificial neural network to determine whether or not a given image includes a fish. The robotic element in the project is provided by a mechanism in which images on pieces of paper are fed into an area where a digital camera can capture the image and use the digital images as input into the ANN. Together with the artificial neural network, this process intends to serve as an automation of the fish counting process in a hydroelectric dam environment in the future.
GoodBoy (the name of the robot) has two behaviors: (1) line follow; (2) recognize a red object based on it's color, and once it recognizes the object it opens it's mouth and barks. GoodBoy uses a Q-learning method to train its line-following feature. The states are a combination of the 3 light sensors. on and off states. The actions are forward, turn slightly to the left, and turn slightly to the right. The positive award is given to the robot if at least one of the sensors stays on the line. The robot gets negative award if none of the sensors are detecting the line. to achieve more complicated behaviors. Currently Natasa is a graduate student at the CS Department. If you have any questions on this Lego robot, you can drop her an e-mail: natasa@boucafe.com
RoachBot learns through reinforcement. It seeks dark areas and hides in them. Response to collisions was added as a tertiary feature, to decrease the likelihood of the robot getting stuck trying to go to a dark spot that was really a wall or other obstacle. The machine learning algorithm chosen for use in RoachBot is is an integration of an artificial neural-network as an approximation of the Q function in order to allow for some abstraction of the importance of various states. The significance of the RoachBot project is that it provides a demonstration of the capabilities of machine learning, as well as an illustration of the difficulties present in using learning algorithms to accomplish what seem to be simple tasks. It also provides insight into the complexities of the behaviors that common insects exhibit naturally. In order to maintain simplicity and focus upon the machine learning aspects of the project, RoachBot was implemented in standard LEGO Mindstorm parts, and uses a LEGO Mindstorm RCX as the brain.
Sam was inspired by the idea "subsumption architecture" [1, 2], in which robot behaviors can be added together to achieve more complicated behaviors. He applied this idea into his artificial neural network Lego robot shown at the left. This robot makes decisions based on a subset of the situations that the robot "sees" as it moves through its environment. This robot has 5 behaviors: pattern searching, building maps, wide exploration, light finding, and local obstacle avoidance. It has two light sensors and two touch sensors wired to the same input, operating as an AND function. Internally, the robot has a set of recurrent neural network which were trained rather than hard coded. The computer language used to program this robot is LejOS. Currently Sam has a job at the Microsoft and work hard to finish his M.Sc. degree at the same time. If you have any questions on his Lego robot, you can drop him an e-mail: samsk@microsoft.com.
Johnny used an artificial neural network as a mapping tool to implement the behavior-oriented Lego robot (shown at the left). The input are the observed environment casted as a vector and the output from the neural networks are casted into the navigation of the robot. Johnny's robot has three behaviors: following a black line, obstacle avoidance, and detecting the edges so that it would not fall into. Johnny also developed a simulation environment for his robot and the training process was done in this simulated environment before down load to the Lego robot. Johnny's robot has four light sensors and four touch sensors. Internally is a 12 X 8 X 11 neural network trained with back propagation. Johnny has graduated Fall 2003 and currently working in Taiwan.
Contact Us | ©2004 Western Washington University