Reminder: Readings are your responsibility. You will be expected to come to class prepared, having read the material, and ready to participate in the discussion
Logistics
- How is Lab time going?
- Tour of Labnotebook2!
Localization
- What exactly is the problem with knowing where you are?
- Remember all of this needs an agreed upon coordinate system!
- How would you do it with your eyes closed?
- Recall odometry is like dead reckoning
- Lots of noise in the signal; accuracy varies; errors build up
What kinds of landmarks can be used?
- Lidar detected fixed obstacles
- Vision detected fixed obstacles
- How can one obstacle be distinguished from another?
- Color, Size, Shape
- Note that in addition to recognizing, we need to know the distance
Lidar detection
- Robot has a ‘map’ of fixed obstacles
- Robot compares that map with the apparent, transient, map from Lidar
- Calculates a probability distribution of where it might be on that map
- Process is called AMCL - Adaptive Monte Carlo Localization
Visual Detection
- Robot has a collection of scenes it can recognize.
- Needs to meet certain characteristics for it to work
- With each scene is the coordinates that correspond to the recognized scene
- CV is constantly analyzing what is seen by the camera
- As soon as it identifies an image, it can use that to figure out where it is
- Note! It has to also figure out where it is relative to the image
Requirements Visual localization
- CV (computer vision) algorithms need to be able to identify and differentiate it from other images
- A coordinate in 3D space is required
- It needs to stay put and not move
- Examples: facade of a building, a particular tree, a wall, etc.
- Bad examples: a parked car; a person
Fiducial “SLAM”
- Special CV Algorithm identifies a known pattern – a fiducial
- Each fiducial has a unique ID and a known 3D pose in the room
- Algorithm determines the TF relationship between camera and a fiducial
- As fiducial has a known 3D pose relative to the room, this gives a 6D pose estimate for the camera
- This is then used to localize the robot relative to the map
- Need to be able to see the fiducial and identify it
- Distance? Size of Fiducial? Is it in frame? Is it occluded?
- Need to see more than one fiducial to get a position fix
AMCL - Important algorithm for localization
- Adaptive Monte Carlo Localization
- Very sophisticated (and standard) mathematical technique
- Can be used with different kinds of sensors.
SLAM and AMCL
How AMCL works with Lidar
- Requires a map of Lidar visible obstacles, with a coordinate system, and anchored in the real world
- Given a stationary robot, and a lidar scan, what does it see?
- Look for a match on the map
- Form a probability distribution of where the robot MIGHT be
- Move the robot a little.
- Compute what the lidar would see given what is known about the motion
- Update the probabilities
SLAM
- What if there is no map yet?
- Simultaneous localization and mapping
- Create a theoretical map based on view of the LIDAR
- Move the robot a little and update the map
- Travel through the relevant region (using e.g. Teleop)
- Use the map to localize, or where there is no map yet, extend the map.
GOAT - Go To Any Thing
- https://theophilegervet.github.io/projects/goat/
- https://github.com/facebookresearch/home-robot
Thank you. Questions?
(random Image from picsum.photos)