This is the main control system used to fly the copter from the position reported by the T265 camera.
Written by Mitchell Cooke ([email protected])
In order to run the project one must first run the camera recording app from the repository T265LiveCamData. This file interfaces with the RealSense SDK and writes the camera data to a single line CSV file. The T265ControlSystem project then reads from this single line CSV file to get the camera's current position and orientation. In order to ensure that the position and orientation reported from the camera is accurate, a short calibration must be performed. This involves moving the device upwards and rotating it back and forth for a second or two until the confidence level of the device becomes 3. A confidence level 3 represents the highest confidence in position. A confidence level of 2 or even 1 are both inaccurate and the reported position of the device when moved can be off greatly. The reason that position and orientation is recieved in the python project through a CSV file is because the RealSense SDK is only available for C/C++ and the python wrapper does not work on ARM based architecture (which the Pi uses). In short, the C program T265LiveCamData reads the position and orientation data from the camera and writes it to a single line CSV file. The python program T265ControlSystem then reads this single line CSV file in order to fly the copter.
The control system of this program is written by Keyvan (current PhD candidate) and Mohammad (graduated). Modifications of filtering have been made to allow the camera to report a smoother position and velocity with less noise while minimizing phase delay below 10ms. The matlab code for analysing and designing filters is in the GitHub repository named T265MatlabAnalysis. There are also several matlab programs for analysing the position reported by the T265 camera, which are contained in the repository named T265MatlabAnalysis. Each of these scripts explain their purpose in the code and some sample data has been included so that the programs can be tested. These matlab scripts can be helpful for understanding the accuracy and capabilities of the camera.
Currently, the position filter uses an order of 5 in order to minimize the phase lag below 10ms and the velocity filter uses an order of 1 (no filtering). Thresholding is done as it is common for there to be large spikes in velocity due to noisy position.
Several html files are generated after the program is run that display flight information.
When flown with this program, the drone is very jumpy in the Z direction and can be oscillatory in the XY plane. Force mapping in the Z direction has been experimented with and has been unsuccessful due to the drone not taking off. Force mapping was tested on both the drone controlled by the T265 camera and the standard control system by the main desktop with Motiv to isolate device dependent issues. In both of these tests, the drone was not able to to take off and the propellors spun fast enouch such that one side of the drone lifted off but with no translation.
It is believed that the instability of the drone is due to the inaccuracies associated with the T265 camera. Camera position reporting can vary by about a centimetre even when the drone is stationary. With vibrations from the motors also causing drift with time, this can cause oscillations in the control system. It was observed that about 15cm of drift was created from a 30 second flight. Another problem from the T265 is that if the reported confidence level of the device is not 3, the reported position can be wildly inaccurate. For example, if the device is moved upwards 30cm while the confidence level is 2, the camera may report that the position is around 5cm. Extensive testing in CamDataAnalysis.m in the T265MatalbAnalysis repository shows the probability of the camera reporting a very inaccurate position to be around 50%. This is the reason fast calibration is needed before a flight to ensure that the confidence level is 3.
Delay with the stop button has also been a problem. It is not understood why, but the motors only stop about a second after the stop button is pressed, which is dangerous and can easily damage the drone. Perhaps there is some form of delay with the serial ports on the Raspberryh Pi, however, there is no delay with the T265 camera as when it is moved the postion updates on screen instantly with no noticeable delay.
It was attempted to see if the instability of flight is due to latency with the Raspberry Pi. This was done by trying to send Motiv motion capture data of position and orientation serially to the Pi. This requires a USB 3.0 male to male connection, which was not able to connect successfully to communicate between the computer running Motiv and the Pi.
The first step with future work is to see if there is a way to successfully implement force mapping in the Z direction. Another way to improve stability of the copter may be through sensor fusion, possibly with GPS if the drone is used outdoors.