15-494/694 Cognitive Robotics:
2017 Final Projects

Optical Flow for Door and Obstacle Detection

• Use textured wallpaper as features.
• Find doors by looking for gaps in feature clusters.
• Use door edges to align robot for perfect travel.
• Use optic flow to detect obstacles without markers.

Tic-Tac-Toe Player: Robot vs. Robot

• Two robots playing tic-tac-toe on a physical board.
• Use vision to detect moves.
• Manipulate physical game pieces.
• https://youtu.be/KWnk6U6LCTM

Tic-Tac-Toe Player: Voice Controlled

• Two robots playing tic-tac-toe on a physical board.
• Humans use speech to choose moves; robots execute them.
• Manipulate physical game pieces.

Object Manipulation: Cubes

• Changing the orientation of a cube with minimal translation.
• Building a wall by tightly lining up cubes.
• Interacting with magnetic cubes (grasp, transport, release).

Object Manipulation: Chips and Balls

• Getting an object away from a wall or corner.
• Separating an object from a pile.
• Capturing a ball with a loop and moving it around.

Maze Navigation Without Markers

• Recognize walls using optic flow, IR reflectance, or detecting wall/floor edges.
• Map the maze.
• Find a goal object in the maze.
• Plan a path to get the object out of the maze, and execute.

Object Learning

• Use the Inception convolutional neural net to learn to recognize objects from Cozmo camera images.
• Detect object presence, not full pose.
• Possibly determine whether the object is in the left or right half of the image, or centered.

Multi-Robot Chip Sorting

• Use color images to classify chips.
• Visually recognize containers.
• Multi-robot coordination to bring robots together.
• Use "forklift" to transport chip to second robot.
• Second robot uses "dustpan" to put chip in container.

Cozmo's Magic Dream House

• Three-story house with working elevator.
• Enter and leave the house.
• Navigate between rooms, and between floors.
• Operate the elevator by pressing a button.
• Move objects around within the house.

Improved Motion Primitives

• Use machine learning to self-calibrate move and turn actions.
• Write an improved algorithm for approaching and picking up light cubes.
• Better failure detection and recovery for cube pick-up.
• Object pushing or tumbling operations.

Cozmo Draws

• Show Cozmo an image and he extracts the major edges, and draws them as a sequence of lines and arcs.
• Uses some image processing code to reduce an image to broad strokes.
• Requires a pen attachment to the lift, and software to control it.

Dave's To-Do List

• Update world map automatically.
• World map viewer.
• Path planning with turn constraints, for chip pushing.
• Color image support.