Chapter 7, Avoiding the Stairs

1. LIDAR, a type of laser radar, is the most common SLAM sensor used, by a wide margin. The 3D data that LIDAR provides is perfect for SLAM’s mapping function.
2. The Wheel odometers reduce the search space that the SLAM algorithm needs to look for possible locations of the robot after moving. Thus, it increases information and reduces uncertainty in the map.

3. It reduces noise and gets rid of stray single pixels in the image, making for a smoother result.
4. Instead of using radial red lines, the program can just draw upwards from the bottom of the screen in a series of vertical lines.
5. We just want to use the upper half of the room to train the network because the lower half has the toys on it and are subject to change. The upper half of the room does not change with the addition of toys.
6. We don’t have to have a map to successfully navigate the room. We are providing labeling of our training set by just driving the robot around and taking pictures at regular intervals. This approach is also far more resilient to changes in the room, like throwing toys all over the place.
7. We would need to have a navigation function that determined where in the room we were at – this would probably mean a SLAM algorithm. We would also need a something to detect the stairs.
8. We trained the robot to navigate by looking at the upper part of the room. We only drove the robot in safe areas, and used that information to allow the robot to predict its next driving command based on where it is in the room. Since we hopefully did not drive the robot down the stairs in this process, the robot will never get a command to drive towards the stairs. We have to be careful to get a good training result before letting the robot loose, however. I used a “baby gate” to block the stairs for early testing.  We can add a lookdown sensor to detect stairs as an additional safety measure. I would use an IRPD (Infrared Proximity Detector) for this purpose.