When developing navigational tools for autonomous robots for tasks like mine clearing and mine detection, GPS is not always an accurate option. One alternative to help circumvent this problem is to use LIDAR. LIDAR, sometimes known as “light detection and ranging”, is used to acquire depth information about a sensor’s surroundings, which, in conjunction with regular visual cameras, can provide an autonomous robot with the information it needs for visual odometry.
LIDAR sensors come in 2D and 3D varieties. While the 2D sensors can not inherently produce a three dimensional map, they can be affixed to a rotating platform that rotates in order to generate three dimensional maps. This type of “nodding” platform can also be used with 3D lidar, allowing for faster and/or more thorough imaging. Pictured below is an example of how a nodding platform can be used to improve the quality of a 3D capable LIDAR.
Here is a clip of one of the LIDAR models undergoing a nodding test:
Scanning while subjecting the LIDAR to a rotational trajectory can greatly increase the resolution of a scan vs a single, stationary scan of the lidar. Pictured below is a comparison for our lab space (top) of a single scan of the lidar (middle) and a nodding scan (below), also called a Raster scan. Note that we have post-processed the LIDAR data to appear similar to a color image, but in our case, the whiter a region, the closer the object is.
It is clear that the latter image has far more information and would be preferable to a single scan. In order to generate an image like this, a reliable cradle is required to rotate the lidar. For that, a controller is required. What follows is a simple example of how feedback control can be used to make such a cradle.
Depicted below is a simple PID control loop. In essence, a physical system, or “plant”, has a desired setpoint. In this case the plant is the cradle and the setpoint is an angle at any given time. The difference between that desired value and the actual value is then treated as the error in the system. That error is then combined with three types of gain, summed, and then sent to the plant. This cycle continues as error is recalculated and the controller tries to correct it. In this example, the error is how far off the angle is from the desired angle, and the control signal is sent to the motor in the plant.
The three gains are proportional, integral, and derivative. Proportional gain is the simplest of the three; it provides a signal that is proportional to the size of the error. In the case of the cradle, when the angle is farther from the setpoint, proportional gain sends a stronger signal to the motor to try and correct the error.
If one were to only use proportional gain the system might never settle on the right location. It could overshoot it and stop, undershoot and stop, or even oscillate indefinitely. That is where integral gain comes into play. The integral gain is multiplied by the total sum of the error each loop. The strength of its response builds when the plant remains at the wrong angle. Integral gain will allow to controller to settle at the right angle when proportional gain alone can not steady the system.
Lastly, there is also a derivative element. High proportional and integrative gains can lead to the plant overshooting its desired value. In order to add some damping to the system, the derivative gain opposes the current motion of the plant more strongly the faster it moves. When all three gains are tuned correctly, the LIDAR cradle moves to its desired setpoint quickly with little or no overshoot and settles on the correct value.
Media Sources:
1.) https://www.google.com/search?q=PID+control&source=lnms&tbm=isch&sa=X&ved=0ahUKEwiRprK5g5HcAhUYO3AKHQKMDJ8Q_AUICigB&biw=1299&bih=675#imgrc=Odx-S5Cmz1QZZM:
2.) Benson et al, DSCC 2018
Recent Comments