Technical Details

Initial Approach

Initial approach used 3 distinct landmarks to localize dog in one of the 3 coves in the pen. The bone must be located next to one of these 3 landmarks.  Then used infrared readings and knowledge of map structure to navigate and look for landmarks, then the bone.

Pros

• Easy to determine and localize which part of pen
• Easy implementation -- basically walking until can no longer move or see target

Cons

• Never really know where exactly on the map you are, only a general idea
• IR readings sometimes flaky, throws off trajectory
• Not smart/interesting

Current Approach

Apriori Information

• Dimensions of world boundaries (thanks to last year student)
• Locations of uniquely sequenced, colored, line landmarks in world
• Specification of which landmarks are viewpoints (IE: can view a section of the pen easily)

Algorithm

1. Dog awakens and looks for 2 different-colored landmarks
2. Particle filter is called to localize
3. Calculate path to closest viewpoint
4. Dog walks along path and localizes at intermediate landmarks
5. When dog reaches viewing landmark and looks for bone:
• If bone is present, fetch it and retrieve to origin location
• If no bone found, try a diferent viewpoint
6. Repeat steps 3-6

System Components

Unique Features

Path Planning

To represent the world a visibility graph is constructed that shows the connection points between different landmarks. A landmark is connected to another landmark by an edge if when the dog is at one landmark it can rotate 360 degrees and see the other landmark. This lets the dog construct a general view of the world, it does not actually know the layout of the pen. Using this graph Djikstra's Algorithm is run to provide a path from the current landmark to the destination landmark. This is necessary because the dog may want to go to a destination landmark that it is not currently connected to in the visibility graph so the path planning provides the small steps in getting there.

Course Related Features

These are the parts of the system divided into the sections that we covered in class. This shows the depth of the project and the amount of the Tekkotsu framework it uses. 

Motion Commands Visual Routines Map Builder State Machines Pilot Kinematics Grasping 

Visual Routines

In order to determine the location of the landmarks and distance from objects visual routines are created for simplicity. The size of landmarks in the shape space determine distance and the proximity of certain visual elements show location as well. 

Map Builder

To populate the shape space the map builder is predominantly used. This is also necessary for the localizing as the particle filter needs a populated world shape space to determine the dog's position 

State Machines

Segragation of code is achieved through the use of state machines. This allows the bone grabbing and world localization to be separated as well as indicates which functions are currently being run. 

Kinematics

The joint values for the head manipulation is achieved using the kinematics engine. Some of the offset values are calculated to achieve the desired link and joint positions. 

Grasping

After the Kinematics and Motion Commands are done to get the dog to the bone a grasping node is fired to actually grasp the bone in the dog's mouth, which shows the interaction between the dog and other objects.