Embedded Motion Control 2019 Group 7

From Control Systems Technology Group

Revision as of 14:23, 20 June 2019 by S154460 (Talk | contribs)
Jump to: navigation, search

Embedded Motion Control 2019 Group 7: PicoBello

Credits to Group 1 2019 for the wiki layout

Group members

Name Student nr.
Guus Bauwens 0958439
Ruben Beumer 0967254
Ainse Kokkelmans 0957735
Johan Kon 0959920
Koen de Vos 0955647

Introduction

This wiki page describes the design process for the software as applied to the PICO robot within the context of the "Embedded Motion Control" course project. The project is comprised of two challenges: the escape room challenge and the hospital challenge. The goal of the escape room challenge is to exit the room autonomously as fast as possible. The goal of the hospital challenge is to autonomously visit an unknown number of cabinets as fast as possible.

Design Document

The design document, describing the initial design requirements, functions, components, specifications and interfaces, can be found here.

Challenge 1: The escape room

As an intermediate assignment, the PICO robot should autonomously escape a room through a door from any arbitrary initial position as fast as possible.

To complete this goal, the following steps are introduced:

  1. The sensors and actuator are initialized.
  2. From the initial position, the surroundings are scanned for gaps between the walls. If no openings are found, the robot is to rotate until the opening is within view.
  3. The robot will drive sideways towards a subtarget (to prevent loosing the target when moving in front of it), placed at a small distance from the opening in order to avoid wall collisions.
  4. Once arrived at the subtarget, the robot will rotate towards the target.
  5. After being aligned with the target, the robot will drive straight trough the corridor.

To this end, methods for the detection of gaps between walls, the placement of (sub)targets and the path planning are required.

Data preprocessing

Beforehand, some of the measured data can already be labelled not useful. That data is summarized below.

  • As the provided data structures have built in minimum and maximum values, data that is out of these bounds is first filtered out. This concerns the laser range finder radius and angle.
  • Data can have outliers or unexpected readings, therefore data is removed using a low pass filter. Datapoints that have no close by neighboring points are therefore removed.
  • Next to the initial angle correction, gap detection only works in the range of > 0 and < 180 degrees in front of PICO. This is to prevent reading out data that is perfectly in alignment with PICO's x-axis (lateral direction). This is to prevent data readings as depicted in the left Figure below.


Gap detection

From the initial position, the surroundings are scanned using a Laser Range Finder (LRF). A gap is detected when there is a significant jump (of more than 0.4 [m]) between two subsequent data points. This method is chosen for its simplicity, it does not require any mapping or detection memory.

Within the context of the escape room challenge, either from within the escape room or from within the corridor a gap can be detected, as depicted in respectively the left and right Figures below.

As can be seen in these Figures, dependent on PICO's orientation, gaps can be detected under an angle resulting in a sub optimal recognition of the corridor entry. To this end, the gap size is minimized by moving the points on the left and right side of the gap respectively counterclockwise (to the left) and clockwise (to the right). The result of this data processing is shown in the left Figure where the dotted orange line depicts the initial gap, and the straight orange line depicts the final gap. Here, the left (red) point is already optimal and the right (blue) point is optimized.

In the right depicted example of a possible situation, PICO is located straight in front of the corridor. In this case the gap cannot be optimized in the same way because moving either of the points increases the distance between the points (as they are not iterated at the same time). However, the detected gap can be used as target for PICO and should result in a successful exit of the escape room if followed correct.y

Dealing With Boundary Scenarios

Initially 2 detected data points are be expected, 1 of each corner of the exit. This would be the case if the escape rooms exit would have empty space behind it. Topology wise this is true, but in practice the LRF will possibly detect the back wall of the room the escaperoom is placed in or other obstructions standing around the exit.

In the case a (more or less solid) backwall is detected, the gap finder will find 2 gaps, one on each side of the exit. In this case the closest left and right points are taken to set the target. The target is defined as the midpoint between these two closest data points.

(Sub)Target Placement

As driving straight to the target, the midpoint of a detected gap, should actually result in a collision with a wall in most circumstances (refer to the left Figure in the previous section), a subtarget is created. The target is interpolated into the escape room to a fixed distance perpendicular to the gap. This point is first targeted.

PICO will turn on its central axis so that it is facing parallel to the direction of the sub target to the target, and will drive in a straight line to reach it (which is therefore lateral). When PICO is approaching the subtarget, the corridor will become visible. As this removes the discontinuity in the data (as PICO can now see into the corridor), a new target is found consisting out of either 2 or 4 points. In the first case no backwall is detected, in the second case a back wall is detected. In this scenario two gaps will be found; to the left and to the right of the end of the exit. The new target will become the middle of the 1st and 4th point as again the closest left and right point are taken. A subtarget is set, but will not change PICO's desired actuation direction.

Path Planning

The loop of finding gaps is gone through continuously. The following cases are build in:

  • 0 data points found

Protocol 1

If no data points are found, PICO should turn until it finds two valid data points which form a gap. If still no data points are found, PICO assumes he is to far away from the exit so switches to protocol 2.

Protocol 2

If after a full turn not data points were found, apparently no gaps are in range. PICO should move forward in a safe direction until it either cannot move anymore, a maximum distance is reached, or data points are detected. If none are detected, return to protocol 1.

  • 2 data points found

The 2 points are interpolated as target, the sub target is interpolated. PICO should align itself with the sub target to target direction and move lateral towards it the lateral error is to big. If the lateral error is small enough, move straight forward.

  • more than 2 data points found

The closest two points are used to calculate the gap.

Dealing With Discontinuities In Walls

If a wall is not placed perfect, the LRF could find an obstruction behind the wall and indicate it as a gap (as the distance between the escape rooms wall and the obstruction will be bigger than the minimum value of a gap), while in practice the gap is too small to fit through and is not intended as the correct exit. These false positives should be filtered.

In the Figure below the algorithm used is depicted. From right to left (as this is the direction PICO stores data) first the actual gap is detected (yellow line), then 2 other gaps are detected as the LRF can see through the wall. The first point of every jump (the blue points) are checked for data points to the left that are too close by in local x and y coordinates. If this is the case, the gap cannot be an exit. Likewise, the left (the red points) are checked for data too close by to the right. This is done over all available valid data in the corresponding direction. If data is detected which is too close by, the set of points which form the gap are deleted. Note that the data will not be deleted when a back wall is detected behind the actual exit as the desired gap to be found is big enough.

This detection algorithm will break the exit finding procedure when a back wall is placed to close to the exit, but it is a given that the escape room will have an open exit.

When the algorithm i this to data points found in an actual gap. These points will never be too close by as the blue point will never have a data point to the left closer by then the minimum exit size (as will the red point to the left). Again (correctly) assuming that the escape room has an open exit. Important to note is that this check is done with the initial gap detection data points, not with the optimized location. If it would, in this example the blue point of the actual exit has directly neighboring data to its left (solid blue circle)!

Challenge Review

We became second with a finishing time of 60 seconds!

Initially 2 doors where introduced in the escape room. PICO would not have been able to find the correct exit as no check was implemented for the extra requirement that the walls of the exit are 3 m. PICO would have driven to the exit it saw first. If both exits would have been seen at the same time, PICO should have chosen the closest one with a risk of constantly switching between both if the distance is similar. The implementation of first aligning to the direction of the exit would have helped as the other gap should not have been visible anymore.

Despite of completing the challenge successfully, some undesired behavior was observed:

  • Due to the long exit walls, PICO kept compensating its orientation position in front of the exit. Buffers which should have limited this behavior where implemented (a valid range for the orientation instead of a single value), but due to the big distance these are currently assumed to be non-sufficient. Also, due to the length of the exit the exit opening was closer to the obstacles around the RoboCup field, which is expected to have partially messed up the non-valid gap detection. Luckily there was a moment where the orientation was correct which moved PICO forwards.
  • PICO stopped very close to the finish line. We expect this happened due to too much disturbance behind the escape rooms exit. We always tested with short exit corridors (as explained above), so during the challenge the exit was located much closer to the table standing in the hallway. We expect some other or even no valid exits were seen for a short moment. Luckily PICO did find the valid exit again and drove over the finish line!
  • During the first try PICO hit the wall, so the safe distance to a wall should be validated. As we do not have the data, we do not know if the collision was caused by overshoot of the initial driving to sub target or due to not reacting/measuring the wall that was too close. As the second try was programmed to turn to opposite direction for gap detection, we probably avoided this error.

Our lessons learned during the Escape room challenge are that we should:

  • be more strict in how we interpreted world information (two exits was technically allowed as one was defined with the correct corridor wall length),
  • not assume that certain scenarios (longer walls) work without testing,
  • validate the collision avoidance.


Challenge 2: Hospital

Overview

For the hospital challenge, the software design can be split into four parts: Perception, Planning, Monitor and Control. These parts will be explained in the sections below. A schematic overview of the structure and coherence of these components can be seen below. Because it was desired to run several interacting processes at the same time, a software design has been made that can cope with this. The implementation of this design is shortly explained in the section Implementation.

Software Architecture Group 7

Perception

The perception component is responsible for translating the output of the sensors, i.e. odometry data and LRF data, into conclusions on the position and enivorment of PICO.

Localization on the global map is performed using a particle filter. Detection of static obstacles is done using a histogram filter.

Monte Carlo Particle Filter (MCPF)

In the hosptial challenge it is very important that PICO is able to localize itself on the global map, which is provided before the start of the challenge. The particle filter is responsible for finding the starting location of PICO when the executable is started. After the initilization procedure it will keep monitoring and updating the position of PICO based on the available LRF and odometry data.

General Implementation

Initialization

When the excecutable is started PICO does not know where it is, besides being somewhere in the predefined starting area. In order to find the most likely starting position (more on this later), an initialization procedure was devised which is able to, whithin a limited time frame, find a set of possible starting positions. Finding this starting position is best described by the following pseudo code.

...
loop for N times
{
Intialize particles randomly
Calculate probability of particles
Pick and store X particles with largest probability
}
...
Initialisation of the particle filter

This algortihm provides the particle filter with a list of particles which describe the possible starting position of PICO. This list of particles is then used as an initial guess of the probability distribution which describes the location of PICO. This distribution is not necessarily unimodal, i.e. it is possible that a certain starting position delivers identical data with respect to another starting location. It is therefore not possible to immediately conclude that the mean of this distribution is the starting location of PICO. The probability distribution stored in the particle filter must first converge to a certain location, before the mean position can be send to the world model. Convergence of the probability distribution is tested by calculating the maximum distance between "likely" particles, in which a "likely" particle is defined as a particle with a probability larger than a certain value. This distance is constantly monitored to determine whether convergance has been achieved.

Often it is not possible for the particle filter to converge without additional information. This is for instance the case when PICO is in a corner of a perfectly square room with no exit and perfectly straight corners. In this case the particle filter will first signal to the state machine that it has not yet converged, the state machine will then conclude that it has to start turning or exploring to provide the particle filter with extra information on the current location of PICO. When sufficient information has been collected to result in an unqiue location and orientation for PICO the particle filter will signal the state machine that it has converged. After convergence the particle filter will start its nominal operation, described below.


Propagating on the basis of odometry data

After initalization of the particle filter the starting position of PICO is known, however PICO is a mobile robot so its position is not constant. It is possible to run the initialisation procedure every iteration to find the new location of PICO, this is however not very efficient. We know the probability distribution of the location of PICO in the last iteration, before it moved to a new location. Additionally we have an estimate of the difference in location since we last calculated the location of PICO, this estimate is namely based on the odometry data of the wheel encoders of PICO. This estimate can be used to propagate all particles, and thereby propagate the probability distribution itself, to the new location.

In simulation this propagating of the particles on the basis of the wheel encoders would exaclty match the actual difference in location of PICO, as there is no noise or slip implemented in the simulator. In reality these effects do occur though. In order to deal with these effects a random value is added to the propagation of each particle. This makes sure that the particles are located in a larger area than would be the case if the particles were propagated naively, without any noise. This larger spread of particles then ensures that the actual change of location of PICO is still within the cloud of particles, which would not be the case when the spread of particles was smaller. The amount of noise that is injected in propagating the particles is a fraction of the maximum allowed speed of PICO, in the final implementation this fraction is set to 10 percent.

In pseudo code the propagating of particles can be described in the following way:

...
for all particles
{
X = sample(uniform distribution (a,b))
Y = sample(uniform distribution (a,b))
O = sample(uniform distribution (c,d))

x location  += xshift + X 
y location  += yshift + Y
orientation += oshift + O
}
...

Find probability of each particle

Up until now the described particle filter is only able to create and propagate particles, however it is not yet able to conclude anything about its current position in any quantative way. A probability model for each particle is needed before any conclusions can be drawn. The LRF data is the perfect candidate to draw conclusions on the current position of PICO. The LRF data does not have the disadvantage of the odometry data, i.e. that it is unreliable over large distances. The LRF data by definition always describes what can be seen from a certain location. It is possible that there are objects within sight, which are not present on the map, however it will later be shown that a particle filter is robust against these deviations from the ideal situation.

The probability of a particle is defined on the basis of an expected measured LRF distance and the actually measured distance. This however assumes that the expected measured LRF distance is already known. This is partly true, as the approximate map and particle location are both known. Calculating the distance between one point, given an orientation and map, is however not a computationally trivial problem. In order to efficiently implement the particle filter, given the time constraints of the EMC course, it was chosen to use a C++ library, to solve the socalled raycasting problem. The documentation, github repository, and accompanying publication of the range_libc library can be found in the references. [1] In order to be able to run the particle filter in realtime the fastest algorithm, i.e. Ray Marching, was chosen to be used in the particle filter. The exact working of these raycasting algorithms will not be discussed on this wiki.

With both the measured and expected distance (based on the particle location) known, it is possible to define a probability density function (PDF), which maps these two values to a likelihood of the considered ray. In this implementation the decision was made to choose a PDF which is combination of a uniform and gaussian distribution.[2] The gaussian distribution is centered around the measured value and has a small enough variance to drastically lower the likelihood of particles that do not describe the measurement, but a large enough variance to deal with noise and small mapping inaccuraries. The uniform distribution is added to prevent measurements of obstacles immediately resulting in a zero likelihood ray. The likelihood of rays is assumed to be independent, this is done in order to be able to easily calculate the likelihood of a particle. When the measurements are independent the likelihood of a praticle is namely the product of the likelihoods of the rays. After determining the likelihood of each particle, their likelihoods need to be normalized such that the sum of all likelihoods is equal to one.

In order to implement the above stated probability model two important deviations from this general setup were made. Firstly it was noticed that it is difficult to represent the particle likelihoods, before normalisation, using the standard C++ data types. Given that the above stated probability model has a maximum value of approximately 0.6, it is easy to compute that even likely particles, which each contain 1000 rays, will have likelihoods smaller than 1e-200. In order to solve this problem another way of storing the likelihoods was devised. For the final implementation a scientific notation object was developed. In this object two values are stored seperatley. A double to store the coefficient and an integer to store the exponent. This allows us to store a large range of very small and very large values.

A second implementational aspect has to do with the maximum achievable excecution speed. One can imagine that the before stated raycasting algorithm is quite heavy to run, compared to other parts of the software. In order to prevent slow downs the particle filter is assigned a separate thread, as discussed before. However to further reduce the excecution time a decision was made to down sample the available data. Out of all 1000 available rays, only 100 are used. There is a risk of missing features when heavy down sampling is used, however no significant reduction in accuracy has been noticed due to down sampling the available data. In literature, most notably the source of the ray casting library, down sampling the LRF is standard practice in order to reduce the computational load.[3]

Finding Location of PICO

With the probability distribution (represented by the particles and their likelihoods) known, it is possible to estimate the location of PICO. Ideally one would take the weighted median of the probability distribution to find the most likely position of PICO. However taking the median of a three dimensional probability distribution is not trivial. One could take the median over each dimension, however this is not guaranteed to result in a location which is actually in the distribution. The initialisation procedure was devised in order to tackle this problem. The procedure makes sure that the distribution is unimodal. Given this unimodal distribution it is possible to take the weighted mean of the distribution to find the most likely location of PICO.

Snapshot captured by PICO at cabinet 1. Prediction (blue) vs Measurement (green). The "Sofa" in the corner of this room is clearly visible in the LRF data


Resampling

The "real trick" [2] and important step of the particle filter algortihm is resampling. A particle filter without resampling would require a lot more particles in order to accuratly describe the earlier discussed probability distribution. This hypothetical particle filter would have a lot of particles in regions where the likelihood is very small. The basic idea of resampling is to remove these particles from the particle filter and replace them either with a random particle, or with a particle placed in a region with a high probability. Selection of particles is done in a non-determenisitic way, a particle is chosen with a probability of its likelihood.

In resampling it is often preferred to also inject random particles in the filter. Random particles allow the particle filter to re-converge to the true location when the current estimate is no longer valid. The risk of random particles is most apparent in an identical room situation. A particle that is initialised in a situation close to identical to the true location will lead to a risk of losing the true estimate. In this implementation random particles are not injected into the filter during normal operation, to prevent losing the accuracy of the estimate.

Resampling is done using a resampling wheel..[4]

Cite error: <ref> tags exist, but no <references/> tag was found
Personal tools