15-494/694 Cognitive Robotics:
2019 Final Project Ideas
Inter-Room Path Planner
- Extend current path planner to generate more complex navigation
plans with explicit doorway passage steps.
- Replan if a doorway is blocked.
- Maneuver effectively in tight quarters.
Cozmo's Magic Dream House
- Navigate between rooms using the inter-room path planner.
- Navigate between floors, using the elevator.
- Modify the world map viewer to display rooms and room names.
- Operate the elevator by pressing a button.
- Move objects around within the house.
- Produce a cool demo.
The Kidnapped Smart Robot Problem
- The "kidnapped robot problem" is when a robot is picked up and moved to a new place, and must
then figure out where it ended up.
- The cozmo-tools particle filter responds to this by randomizing
the particles and trying to use landmarks to re-localize. But what
if no landmarks are visible in the camera image?
- This project will develop smart strategies for searching for landmarks to help the robot
self-localize. This will involve moving both the head and the body.
- The problem is more difficult when the robot is in a tight space because collisions
with walls or objects can throw off its odometry.
Neural Net Line Follower
- Train a set of neural networks to allow the robot to follow lines on the floor made from colored tape.
- Basic network follows a straight or gently curving line.
- Another network detects forks in the road.
- Another network guides the robot to take the left fork (or the right fork).
Transfer Learning: Gesture Recognition
- Use a pre-trained deep neural network as the hidden layer for a new, rapid feature learner.
- See gesture recognition demo at the Teachable Machine.
- Teach Cozmo to recognize hand gestures using the GPU.
- Links: Teachable Machine source,
and news blurb.
- The original version of Teachable Machine
used SqueezeNet.
The current version appears to
use MobileNet.
CIFAR-10 on Cozmo
- The CIFAR-10 dataset contains 6,000 images of each of 10 object classes (cars, planes, birds, etc.)
- Deep learning models have achieved up to 96.5% accuracy on a separate test set.
- PyTorch includes the CIFAR-10 dataset. Implement a deep learning model on Cozmo and use it to
allow Cozmo to recognize novel images in the learned class.
Forklift Attachment
- Write code to visually detect and dock with a pallet using
the Hexnub
lifting kit attachment.
- Pallet detection could be done with a convolutional neural network.
- Modify the path planner to model the shape of the robot with the lift attachment; this is necessary for
accurate collision detection in the RRT search algorithm.
AruCo and Custom Marker Detector
- Similar to the cube detector exercise, but look for AruCo and
custom markers that are partiall out of the camera frame.
Vector Robot Support
- Vector is a
more advanced robot than Cozmo, with a similar SDK.
- Vector includes a higher resolution camera with a wider field of view, a laser rangefinger, a capacitive touch sensor
on the head, and a microphone array.
- Modify the cozmo-tools package to support Vector.
Fun With Quboids
- Quboids are cardboard cubes with custom markers on the faces and magnets inside.
- Design lift attachment for capturing and dragging quboids.
- Assemble quiboid structures using the magnets to snap them together.
- See
the 2017
projects for an initial take on this idea by David Kyle; there
is room for refinement and extension.
Chip Manipulation
- Chips are the size of quarters or poker chips and can be pushed around with a lift attachment.
- Since we can only push, not pull them, the path planner must be restricted to make only shallow turns.
- The current path planner has code for this that is not completely correct.
- It would be useful to design other chip manipulation operations, such as flipping a chip.
|