DJI Engineer Deeply Analysis Phantom 4 Innovation Technology

(Last Updated On: December 24, 2016)[ This Writer is an Engineer who working in DJI] Just as a man attending the whole R&D process of the Phantom 4,  can answer this question “What is the Phantom 4 Innovation Point?”. Have technical introduction, some flight experience. Phantom 4 is a cross age thing, in order to let outsiders to understand, we might talk from the original “what is a smart UAV?”.

DJI Phantom 4 1

What is Unmanned Aero Vehicle/Drone, Micro Aero Vehicle/Drone?

From wiki: An unmanned aerial vehicle (UAV), commonly known as a drone, as an unmanned aircraft system (UAS), and also referred by several other names, is an aircraft without a human pilot aboard. The flight of UAVs may be controlled with various kinds of autonomy : either by a given degree of remote control from an operator, located on the ground or in another vehicle, or fully autonomously, by onboard computers.

Someone think DJI Phantom is not the UAV, it is the flight model. From the academic perspective, this point is wrong. Whether the drones have to be the murder weapon like Raptor? First of all, UAV is no business with the size, see whether there is a certain sense of autonomous flight. Really want to get that whether Phantom is an UAV or not, I suggest to try it in the windy days of ten meters per second.

As for the long criticized that Phantom does not provide ground stations,  so not UAVs, etc…, in fact, you want all the features, including more than one hundred kinds of combined features of the ground station, are in DJI Mobile SDK. But for common users, open so much features will cause danger. Secondly, someone said multi axis can not be weapon, here is a photo from internet:

DJI S900 1

DJI S900 2

The modification of the aircraft using DJI S900. UAV flight control system using A2 control. This modification must be genuine goods at a fair price, although the flight cartridge is PS.
According to my estimate, suspected mount rifle M4 carbine. Although this thing shooting make the aircraft shaking like a dog, but the front fire suppression (scary) is very useful.

DJI S900 3

Although I also hung guns on the Phantom 4, almost was the first people to use Phantom 4 to launch the “shells” in the world. With fireworks, there was exploding, P4 ( Phantom 4) is not damaged, just stick a little ash, indicating that P4 is still very strong.

But in a flagrant way like the above figure ( and the aircraft suspected to use our flight controller and power system), also shook me at the first time. Put the figure just to tell you, don’t underestimate the potential of Phantom.

DJI Phantom 4 13

DJI Phantom 4 14

(This is the first P4+ fireworks shot….. Please don’t die, flight safety) I am not here to discuss what is the aircraft, we can take a look at the first example of UAV, this UAV is one of the earliest time, although everybody says it is flying bomb in the beginning. But do you believe that not a drone which can automatic navigation flight with wings?

UAV earliest 1

Well said so much, just to illustrate that UAV is a very big concept, from the size of the thumb insect aircraft

UAV smallest 1

To the unmanned airship, Rainbow Series

UAV airship 1

Compared to the traditional aircraft, modern UAV is emphasis on intelligence. It is more like a robot instead of ordinary plane, whether independent landing, target tracking, even in search of disaster, all cannot do without the presence of robot technology. I used sense-planning-control three steps which be commonly in this robot technology to talk about the key technologies of UAV. Also mentioned some communication technology.


For a early UAV, in fact, is more concerned about the Sense step. Because it is very weak about the computing ability and sensor features. And in sense inside, the most basic thing is, tell the UAV, where I am, where should I go (but in fact do not let him know who he is, we are not making Skynet).


UAV’s first major event is location, determine where they are, determine their attitude, the early UAV (typical, V1 missile) using mechanical gyroscopes.


Is this kind of stuff to get their own attitude. As for the position, is the use of a simple encoder integral. Such as the V1 rocket is using Pitot tube data points, to determine their fly distance, fly to the right place then head down. Of course, modern UAVs do not use such a primitive means, imagine if a large wind, pitot tube integral, well V1 will be exploded in France, not Britain.

Modern UAVs are equipped with a variety of gyroscopes, such as laser gyro, fiber optic gyro, or we are using MEMS gyroscope. GPS system for UAV provides high convenience. All in the sale of the UAV on the market except Phantom 4 are mainly using GPS and inertial navigation positioning. Optical flow only for speed, it is difficult to give a precise position.

GPS can provide a positioning accuracy of 2.5m for UAV, but you will never see the UAV hovering dangling in the range of 2.5 meters, this is because the modern UAV use the Calman filter, the precision of relative positioning can get higher. We’ve done a laboratory to subtraction of two M100 positioning data, the noise data basically no more than 10 cm. Of course, GPS can also be extended through advanced RTK-GPS positioning accuracy, which is another topic.

Phantom 3 and other DJI, can hover is actually made a compromise on the location without GPS. Take the optical flow algorithm to control the speed. If there is no optical flow, no GPS, control of the UAV is acceleration, this is the difference between the two derivative of control (acceleration) and one derivative (speed). Obviously, to hover, inhibit the speed to zero is much better, this will make the cumulative error integral position as small as possible.

Phantom 4 up to a higher level, VO (visual odometry) positioning. No GPS the flight view control measurement is the position. So precision than Phantom 3 and all the other UAV which sell good in the market. Here to give a very simple example ~

DJI Phantom 4 15

I put off Phantom 4 to 30 cm in height, risking my life, dragging his foot to run a certain distance, P4 will run back. It is simply impossible to finish this if common optical flow algorithm in the interior.

Whether visual observation, or use the GPS and IMU data fusion, data fusion should be carried out through various filtering algorithm. If the UAV has key technology, here is one of the core. Visual positioning is not actually so simple. Visual positioning is one of the SLAM (real-time localization and map building), this should be one of the most core technology of UAV.

SLAM in Prometheus

This figure is the research institutions used to science what is the iconic SLAM map,  Prometheus film screenshot.

Interested friends can go to see the film, Phantom 4 inside the detector with similar algorithm running in the movie, and certainly not as well as the movie, even the current algorithm is not a complete SLAM. But think about you now spend less than two thousand dollars can buy what the movie inside, you will feel I have live to see this.

And P4 inside the active obstacle avoidance function is a very typical SLAM weak application. UAVs only need to know where obstacles, you can conduct Planning, and bypass obstacles. Of course, SLAM can do more than these, including disaster relief, human cooperation with aircraft, even cluster. All about UAV dreams are built on top of SLAM, which is the UAV can fly (after having positioning, attitude determination) era, the core technology of UAV. Is also the difference between modern UAVs and toys.

DJI Phantom 4 16

Although the active avoidance is not very mature, or sometimes hit into the woods (It is difficult to hit the barrier if passive avoidance), but this is a very exciting thing for the first time to show the face itself. I think the Phantom 4 should be listed first in the introduction of mapping and planning UAV products, and maybe be one of the earliest introduction of mapping consumer electronics (here, actually, MI Robot Vacuum of Xiaomi support SLAM mapping function ).

About sense basically here to almost, and there are some infrared cameras, synthetic aperture radar, HD 4K camera, etc.. These are either the task load, is not involved in the UAV work loop, or is no longer a superficial reform. Of course, there is also a method of cheating location, such as Intel has done, is Vicon, Vicon is generally used for small UAV experiments, used to make products belong to curry favour by claptrap.


Another is to talk about Planning, generally speaking the UAV technology, few people mention Planning, or directly to think Planning is a part of Control, it is not accurate. In fact, planning has a very beautiful things and results. Such as a lot of people have seen a TED video lecture: trajectory of many quadcopter. This technology of UAV is Planning. Planning is actually a piece of my love, because the control of small UAV basically has been perfect, needs to be improved is planning.

Like the trajectory, the truth is actually very simple. Draw a line in the physical space, and optimize it according to the index which you want, such as optimization of four order derivative minimum, then you can get a regressed donghonghong curve, flying over the line. Of course, Planning far more than these, for example, now VTOL very hot ah.


How to make the UAV between take-off and horizontal flight conversion perfect? How to quickly stop hovering? This is a typical planning problem. Such as a lot of obstacles, how fast and safe to bypass these obstacles? Is also a very typical planning problem. Even how to get a glider to land on a rope.

planning 1

( From: )

Planning applications are more widely used in Humanoid Robotics, for example, the step control. P4 inside planning, have said before, I will not go into. In my limited experience, the majority of practitioners, and even the University of automation professional, focused on the part of the final module:


I pushed the formula in the control of the pit stumbled to find the north when an engineer one word wake me up: the robot is not the same as the control. Your good control is nothing if your planning and sensing like as dog. This view is also what I want to say, for multi rotor UAVs, control is relatively simple part of the sense (most of the time planning was ignored).

If you have good observation, control is really very simple. A lot of fancier and small businesses, the main content is control all days. Especially many take flight control using STM32 module, again endure STM32 completely not reliable attitude fusion module, and spend a lot of time to study control, even a lot of UAV media is also a large number of advocating the difficulty of control. This may be true for helicopters or fixed wing aircraft, or the Rainbow series you have been saying, but conventional fixed wing controls are simpler than multi rotors. Tracking trajectory is relatively complex. Now more complex control is carried out in planning body, but the use is not too much.

But there is indeed a kind of algorithm we are using, but feel reluctantly if said it is the key technology, because this algorithm is not really complicated, that is artifact, PID! But in fact, just PID is not so accurate, because now most of the use of small UAV attitude control (such as Pixhawk) is actually a nonlinear PID algorithm. Without the speed of planning, position control algorithm is generally PID.

PID controller can use a line to express:

control amount= error *p + error rate * I + error integral * I;

PID controller you can understand as a damped spring will have a cumulative error with time given force. PID is a well-established algorithm, life is much longer than UAVs and even aircraft. For example, a very classic example is Watt to improve the steam engine in the important steam engine governor. This is a very early PID controller, using mechanical implementation.

Small UAV is commonly used in a method based on Li Qun SO3 control method, so-called SO3 is a rotation group. The attitude representation corresponding to the SO3 space is quaternion. If you want to do in-depth understanding, can read some Wikipedias. We use quaternion q1 and the attitude of q0 can generate a relative number of q_relative corresponding to the quaternion, from the current attitude to pose what XYZ axis and big angle turn in the past. Relatively simple to understand, this is the relative rotation corresponding to the current natural angular velocity WX WY and WZ coordinate values, a deep explanation is that you get a relative rotation into tangent space (small SO3) on the map.

quaternion 1

Then at this time the mapping as a control amount thrown to the PID controller, or the angular velocity as a separate level, or simply do another PID controller, you can complete the task of attitude control. Of course, in reality will be slightly more complex than this, for example, in the Pixhawk code inside is to make the yaw rotation as small as possible. But the essence is similar.

This method I have tried in multi rotor, fixed wing (although in the simulator inside), and the work is good. After controlling the attitude, if the multi rotor, helicopter, only the attitude and acceleration corresponding (according to the assumption of Figure), you can complete the speed / position controller. If it is fixed wing, you need to use L1 tracking algorithm to control trajectory tracking. Of course, the other controller is actually a lot, such as the legendary Hinf method, or is said to be a very common fixed wing L1 adaptive Control ( not the same as above L1 , come from a Chinese and the Russia lady). I am also learning.


Communication was not my expertise, and the difficulty of communication technology is less than the current control difficulty, I will not to say much, instead to say some image transmission. UAV a pretty technical content is high-definition image transmission, this is not a lot of UAVs to get.

The main indicators of HD image transmission are delay and resolution. In general, high-definition digital image transmission works: camera data ->H264 compression -> wireless link protocol -> encapsulation protocol- > H264 decompression -> displays. Of course, in order to reduce the delay, should pay more attention to compression and decompression. It needs SoC technology, very difficulty for small factory. Currently DJI is indeed the best in terms of image transmission. In fact even Taobao goods want to achieve its image transmission performance can also buy a half Phantom….


Also have to mention is the Pan-Tilt. Pan-Tilt itself is not so difficult, but the integration is really a very time-consuming and labor-intensive work. Because you slightly change things a little bit, will cause the distance amplification. And several shock absorber quality control is very troublesome. Of course, as a user, I do the earliest aerial four axis with Samsung mobile phone, shake like as a dog, now Phantom are so good to cry.

Pan Tilt


Phantom 4 use VO (visual odometry) positioning, typical SLAM in obstacle avoidance function, Completely planning, Perfect Control,  IT is a cross age UAV.

Want to try? Check it on


Leave a Reply

Your email address will not be published. Required fields are marked *