Reaction time measurement is a process of acquiring timespan for how long it takes for the tested device to change its state after clicking on action element. The most common scenario is the measuring of time needed to load new screen after clicking on the button that invokes screen change. This measurement directly testifies about user experience with the tested system.

So how does our robotic system do it?

The algorithm of reaction time measurement is based on calculation of pixel-wise differences between two consecutive frames. That simply subtracts values of pixels of one image from another. Let’s have two frames labeled as fr1 and fr2, there are three main types of difference computation:

  • Changes in fr1 according to fr2: diff = fr1 – fr2
  • Changes in fr2 according to fr1: diff = fr2 – fr1
  • Changes in both directions: diff = | fr1 – fr2 |

The last mentioned computation of differences is called Absolute difference and is used in our algorithm for reaction time measurements. In general, the input frames are grayscale 8-bit images of the same size. Computation of differences for color images is possible however it would only deliver more errors in RGB color spectrum due to more variables being dependent on surrounding lighting conditions. The final computed difference is just a number indicating the amount of changes between two frames, therefore, it is perfect for detecting the change of screen in the sequence of images.

Enough of theory, let’s make it work!

First of all, we need two images indicating the change of screen. For this purpose, I choose the following two pictures.

Imagine those are the two consecutive frames in the sequence of all frames we talked about earlier.

Next step is to change them into grayscale and make sure they are of the same height and width.

After those necessary adjustments, we are ready to calculate the difference between them. As was told before, it is computed as an absolute subtraction of one image from another. In about every computer vision library there is a method implemented for this purpose, i.e. in opencv it is called AbsDiff. The following image shows the result of subtraction of two images above. As you can see there is a visible representation of both images. That is completely fine because every non zero pixel tells us how much given pixel is distinct from the same pixel in the second image. If result image would be black it would mean that difference is zero and images are identical, vice versa for white image.

Next step is to sum values of all pixels in the result image. Remember that each pixel value for grayscale 8-bit image is in a range from 0 to 255. The difference for this image:

diff = 68964041

This value itself is not very descriptive of the change between the two images, therefore normalization needs to be applied. The form of normalization we use is to transform that computed difference number into a percentual representation of change in the screen with a defined threshold. Threshold specifies what value of a pixel is high enough to be classified as changed so rather than computing sum of all pixels in result image we find how many of pixels are above the defined threshold. The normalized difference for this image:

diffnormed =96.714% (with threshold =10)

This result compared to the previous one very precisely tells us how much change happened between the two images. The algorithm to detect the amount of change between two images was just the first part of the whole time measurement process. In our robotic system, we have implemented two modes of reaction measurement, Forward and Backward reaction time evaluation.

 

Forward reaction time evaluation

Forward RTE is based on real-time-ish evaluation meaning that algorithm procedurally obtains data from an image source and process them as they arrive. The algorithm does not have the ambition to find desired screen immediately but it rather searches for screen changes, evaluates them and then compares them to desired one.

Forward RTE diagram shows the process flow of the algorithm. At the start, it sets the first frame as the reference image. Differences against this reference are then computed with incoming frames. If the computed difference is above the threshold then the frame is identified and the result is compared to desired screen. If this does not match, the frame is set as new reference and differences are then calculated against it. If that does match then timestamp of image acquirement is saved and the algorithm ends. In theory, every screen change during measuring is identified only once, however it strongly depends on the threshold value that user needs to set. Even if this algorithm tries to be real-time the identification algorithms take so much time that it is not possible yet.

 

 

Backward reaction time evaluation

Backward RTE works pretty much the other way. Rather than searching desired image from the start, it waits for all images to be acquired, identifies the last frame and sets it as a reference, after that looks for the first appearance of given reference in sequence.

Backward RTE diagram shows the process flow of the algorithm. First of all, it waits for all frames from the subsequence of frames. After all frames are acquired, the last frame is identified and if the last frame is the actual desired screen then the reference is set and the algorithm proceeds. If the last frame was not desired screen it would mean that desired screen did not load yet or some other error happened. For this case, algorithm records backup sequences to provide additional consecutive frames. If there is no desired screen in those sequences then the algorithm is aborted.

After reference is set the actual algorithm starts. It looks for the first frame which is very similar to the reference one using difference algorithms described earlier. Found image is identified and compared to desired. If identified and desired screens match then the time of acquirement is saved and the algorithm ends. However, if they do not match then the sequence is shortened to start at the index of falsely identified frame and then the algorithm searches further. The furthest index is the actual end of sequence because image at the end sequence was at the start of RTE identified as desired.

Summary

This article contains information about difference based measurement of reaction time. It guides you through computation of differences between two images. Also, this article describes our very own two reaction time evaluation modes which we use in practice.

 

Additional notes

To keep the description of algorithms as readable as possible few adjustments are missing. Preprocessing of the images is the essential part where elimination of noise has high impact on the stability of the whole algorithm. We have also implemented few optimization procedures that reduce the amount of data that needs to be processed, e.i. bisection.

 

Members of our Robot team have participated in the worldwide competition of robots on the 11 April to 12 April in 2015, which was held in Vienna, Austria. There were a lot of competition categories including Humanoid sumo, where we participated. Over 600 robots were registered over all categories and 16 robots in the Humanoid sumo.

Robot specifications

Each humanoid robot has to meet certain specifications, e.g. maximum dimensions and weight. The rules also require having a head, two legs, two arms and a name. We named our robot YSoft Ragnarök, which is a great foretold battle from Norse mythology. The limit for the weight is 3000 g, which was quite problematic for us. At the beginning of the competition we had to reduce robot’s weight to 2997 g by removing some insignificant parts. Our strategy is a great stability which many other robots lack, but it comes with demand for heavy parts especially at the bottom of the robot. Heavy body also reduces mobility and speed so we had to develop better solution in order to find the opposing robot reliably, then move directly to it and wreck it. For this purpose we used ultrasonic sensors with maximum range of 2 meters, high precision and low power consumption.

robot clanekArena rules

Tournament started with qualifications and continued with single match elimination. Each match startswith two robots in opposing corners facing each other. The main goal is to push the other robot out of the arena or to knock it down. If any robot is pushed out of the arena, it can be placed within the arena again, however it must be placed face down. If the robot can autonomously stand up, the match continues. Team gains 3 points for pushing the other robot out of the arena. If any robot falls in the arena, the opposing team gains 1 point. Two points are awarded for a robot that knocks the opponent to the ground. Match ends if any robot is knocked out (and cannot stand up), it does not move for a period of time or the time for the match runs out (the maximum time is 3 minutes). The competitor with highest score wins.

Our performance

We have beaten every robot in the qualification group and successfully advanced into the semifinals without suffering a loss. However, this had already happened in the past when we lost the next two matches to finish in the 4th place. This time we tried not to repeat such outcome. The first semifinal match against Mexican robot Speedy Gonzales was quite even as the opponent avoided contact with us so we only managed to knock it down once. The second match versus another Mexican robot Atom was more one-sided because it could not get up after knockout (this match can be viewed here).

Once we got into the finals, we have faced our old rival from Poland, robot DUE. At the end we have beaten them and won the first place (video). Polish robots DUE and UNO took both 2nd and 3rd place.

 

In QA we use robotic arm to autonomously operate a multifunctional device (MFD) according to a given test that is repetitive, time consuming or not performable by a human.

How does the robot know where is the screen located? How do the 2D screen coordinates transform to the robot’s system?

calib1 clanekRobot moves the end effector (stylus) in the 3D Cartesian coordinate system with the origin [0,0,0] at the center of the first servomotor. Position of the stylus is then transformed to angles of all servomotors by inverse kinematics. It is also possible to calculate position of stylus given the angles of the servomotors (forward kinematics). However the screen of the MFD is a 2D plane in the 3D space with unknown origin, dimensions and rotations. Robot needs to know where the screen is located relatively to its origin in order to correctly tap any button on the screen.

Dimensions of the screen need to be measured by hand in millimeters. We use 2D coordinates with origin at the bottom left corner to define position on the screen. In 3D space position and rotation of any plane is uniquely determined by three non-collinear points. If these three points are known, transformation matrix can be found. This matrix multiplied by position on the plane is equal to corresponding position in 3D space.
Previously we used ‘basic’ calibration where the robot is navigated through the bottom left corner [0, 0] (origin of the screen), top left corner [0, height] and top right corner [width, height]. At each corner, stylus’ position (in 3D space) is saved and transformation matrix is calculated. This method of calibration requires a lot of precision because even a slight deviation from the corner leads to a similar deviation in every robot’s tap so robot might not accurately tap the desired button. There is no feedback from the MFD, but sometimes there is no other way to perform the calibration.

calib2 clanek

With the new semi-automatic calibration we created our custom version of Terminal server (component which handles communication with MFD) that detects any tap on the MFD screen and sends its coordinates (X,Y in pixels) into the Robot application. Screen resolution is also required so Terminal Server sends that on demand. With the knowledge of screen dimensions and screen resolution robot is able to calculate position of a tap (in pixels) to position on the screen (in millimeters) and save the end effector’s position. The semi-automatic calibration procedure is almost the same as the basic one but robot can be navigated to any point within a marked rectangle, not just the specific point at the corner. This nullifies the need for precision. However, in this case a problem occurred in form of inaccurate values of Z axis. For this purpose we have developed an automatic recalibration. This recalibration takes data gained from semi-automatic calibration and automatically repeats the procedure of semi-automatic calibration with the knowledge of existing corners of the screen. It goes through those three corners same as before, however it starts higher above each point and slowly descends to accurately measure the Z coordinate. After recalibration all data from semi-automatic are forgotten and replaced with the values from automatic calibration. This procedure eliminates any error made by an engineer during calibration and makes the robot´s calibration nearly perfect.