John Plewe
CS 154 Robotics Lab
Click here for previous lab


Localization 1 - Constructing "Click"

Introduction

In this lab I will be modifying and extending my extinguisher robot into a new robot dubbed "Click." Click's purpose will be navigation and localization in an office-like environment. He will be equipped with two sonar range-finders atop a servo motor which will do the bulk of the work, as well as bump sensors and possibly light sensors or IR range-finders if necessary. My goal for this lab will be to get Click working, which involves adding the new elements to the robot, getting the sonars to give meaningful input, learning how to use the servo motor successfully, and assembling and using the compass. I will probably create a set of simple test goals for Click to perform in order to demonstrate progress in these areas.


Background

The general problem of the extinguisher project, while not excessively difficult, was an important one, as search and navigation are central problems of robotics. In particular the problem of finding and extinguishing a flame has been solved many times before by many people. But it was ideal for getting my robot built and tested, and did offer several challenges. (Not to mention that my robot still doesn't handle all situations perfectly.) The main problems to be solved are maze searching, light following, and figuring out how to blow out the candle, as well as how to know when you've blown it out. The problem is limited in many ways including that it is in a very simple, enclosed environment, that the light and dark values are known ahead of time, that the height of the candle is known, and that there is only one candle to find and blow out. All of these make it easier to solve the problem with the given hardware.

The localization project extends and focuses upon the searching aspect of the extinguisher project. The central problem I will be attacking is how, given a representation of the environment (a map) and real-time sensor input, can the robot determine where it is. Once the robot is constructed and working, the next lab will be implementing Markov localization to achieve this goal.


Approach


Original Extinguisher Proposal
I intend to implement a subsumption-based architecture (as described by Rodney Brooks in his paper "Achieving Artificial Intelligence Through Building Robots" at M.I.T. in 1986) using a HandyBoard-controlled lego robot. The robot will at least have bump and light sensors which it can use to traverse the maze. As I am using this project as a jumping-off point for future projects, I may also build sonar and compass sensors into the robot, and make use of them if I have time.While I have not yet worked out the details, I envision the following behaviors implemented as a subsumption architecture:
  • Avoid obstacles (respond to bump sensors)
  • Wander (search in darkness)
  • Go towards light (when light is detected)
  • Enable fan (when light reaches certain threshold, until light decreases significantly)
Final Extinguisher Approach
In practice, my implementation turned out to be very similar to this. My architecture had the four following levels, in order of highest priority to lowest priority:
  • Extinguish: Kill all other processes, because we've reached the goal and don't want to be distracted. Slow down motors, and turn on fan until the light decreases to predefined threshold for darkness. At that point, turn off fan and motors. If light is encountered again, repeat this process.
  • Escape: Respond to bump sensors by backing up, turning a random amount, and going forward again. If "stuck" (same command repeated too many times,) back up for longer and spin to a new heading.
  • Follow: If light is above threshold for darkness, follow the light by setting motors to full forward and turning towards light.
  • Cruise: If light is below threshold for darkness, simply go forward. If light is encountered or a bump sensor is triggered, one of the above modes will take over automatically.

Ideally, I believe the robot should spend much of its time in cruise mode, which is basically navigating the maze without any light information. I can simulate this by putting the darkness threshold higher, allowing for the robot to think it is in total darkness when there is some light. But the light provides valuable information and naturally tends to lead the robot to correct paths, so the robot does well in follow mode. Light coming over the top of an obstacle, however, can be dangerous and cause the robot to get stuck against the obstacle, though the random escape routines should be enough to get it out eventually. (The robot's candle light and darkness thresholds can be set interactively when it is started up, by pressing either of the touch sensors rather than the start button and then following instructions.) Any false alarms as to the location of the candle are quite dangerous, as the robot kills the follow, cruise, and escape routines in order to focus on putting out the fire. This was necessary to make sure the fire is put out. If the robot mistakenly stops too soon, and the flame comes back, it just starts up the fan again and keeps working on it.

Just getting the robot built and functioning took a little bit of engineering. I followed a tutorial step by step for the basis of the robot, but then added my own fan tower. My solution for attaching the fan (twist ties) is probably not the best choice, but seems to work despite the awful racket the vibrating fan makes. It took a lot of testing to find light sensors which reported similar values. I had done some of the necessary coding for a previous light-finding assignment in a different course, but had to rewrite much of it as that robot was not successful at several of its tasks. I rewrote the escape routines entirely. In addition, the extinguisher portion was all new to me.

As I mentioned above, the user has the option of providing the robot with new light and dark threshold values interactively. I made this design choice in order to allow for a higher degree of precision and adaptability to different environments. The dark threshold determines how much light the robot must see before it switches from cruise mode to light following mode. The light threshold determines how much light the robot must see to be convinced that it found the candle. It surmises that it blew the candle out when light drops back to the dark threshold. I could have opted for a relative approach, e.g. turn off the fan if the light drops by x amount, but that seemed like it would require a lot more tweaking of the code to get it to work properly. I wanted to be able to take the robot to a new environment, set it in the dark and get the dark threshold, set it next to the candle to get the light threshold, and then have it work for that environment.


Original Localization 1 Proposal
I have a pretty good base to work off of from my extinguisher robot, Trill. I added a pretty solid tower to hold the fan and light sensors, and so it shouldn't be too difficult to mount the sonar unit on top. The unit, consisting of two sonars mounted on a servo motor, was constructed by a previous lab group. I need to figure out how to get it working accurately and reliably, which may or may not be a difficult task. I'll start by trying to get good readings from the sonars, and then figure out the servo motor. I'll also need to add sensors to the wheels, to get some sort of rough estimate of odometry. By the end of the lab the robot should be able to perform some basic functions which involve all of these sensors, such as orienting itself in the midpoint between two walls.

Final Localization 1 Approach
A couple of changes, mostly simplifications, to the design. After looking more into what the other lab had done with the sonar tower, particularly the impressive work they did to get the handyboard to work with two sonars, I decided that a simpler, more solid approach is ideal for my situation. The servo can easily rotate 180 degrees, which means that just one of the two sonars is necessary to look in three of the four necessary directions from the robot without physically turning the robot. My new approach is to use an IR rangefinder for the final direction. This IR rangefinder will look directly to the right of the robot, and the robot will use it to closely hug the wall ont he right side. The sonar can then look left of the robot, forward, and backwards as needed. A quick dropoff due to a doorway will hopefully be differentiable by the IR rangefinder, which an detect up to 80 cm away. It will also be easy to keep the robot going forward, since presumably the wall is straight and it is trying to stay a set distance away from that wall (unless a sudden dropoff above some threshold occurs.) Rotation sensors may not even be necessary.

I've been able to get good results from the sonar. It measures distances pretty much as accurately as my tape measure, and is fairly straightforward to use. The servo is also working well. It is now capable of switching to predefined forward, sideways, and backwards positions quickly and reliably. The sonar will most of the time be checking to the left for doors, and trying to maintain a constant distance from the left wall as the IR rangefinder is with the right. The bump sensors on the front of the robot are also of course working well. The IR rangefinder is more problematic. I'm using a Sharp GP2D02. Frankly, I'm having trouble getting it to work at all, leading to nagging suspicions that it might in fact be dead. More likley, though, it's a problem of hooking it up correctly. I was following this guide and using the provided code, but had little luck. It refers to port PA5, which is also T03. This port seems to be trapped under the lcd, requiring a shunt to get to. (This guide uses T03 to hook up a sonar.) I tried to get a hold of another one, but the others I borrowed were of a different type (Sharp GP2D12). These analog IR rangefinders are less accurate, and require a physical modification to the handyboard to work at all. I'm going to continue looking into the issue and try modifying the code to work with a different timer than T03 while at the same time checking to see if any other IR rangefinders are available. But without it, I can simply switch the sonar to check forward, left, and right, since behind probably doesn't provide very much information, and ideally would provide none after the first move.

Original Localization 2 Proposal
Now that Click, albeit a simplified version due to hardware problems, is built, it's time to implement Markov localization. As per suggestion, I will divide this up into the following 5 steps:

1) Using sonar to follow walls.
2) Creating an environment for testing.
3) Mapping the environment by hand.
4) Performing Markov localization by updating probabilities on the map
5) Determining and printing the highest-probability location, and a list of all probabilities on demand.

Note: I've decided to put my more detailed reports for Localization 2 in the Progress section below.

Progress

Extinguisher
2/4/2004
-
Progress has been severely limited this week as I have been finishing up course work from the previous semester. I have acquired all of the parts and pieces from multiple sources, setup the Interactive C environment on my computer, and intend to have the basic robot constructed by Friday.

2/13/2004 - Robot is built and up and running, with basic light-following, maze-searching behavior. Still doesn't have fan or code for controlling the fan.

2/19/2004 - It's alive! I'm content with the performance of the robot. Fan and everything seem to be working well. Updating final report today. Pictures and movies will be available here by Friday evening.

Localization 1
2/25/04 - Updated site to begin new project. I've acquired the necessary parts, and begun construction. I also updated the old site with pictures, a movie, and source code.

2/25/04 - A little research into the reliability and function of the compass, including this FAQ with its creator, makes me doubt its usefulness for my project. I think I'd better focus on the sonars and wheel rotation info to get an idea of orientation, unfortunately.

3/12/04 - Plans have changed a bit after looking into the sonar setup in more detail, see the Final Localization 1 Approach section for more information.

Localization 2
4/15/04 - Updated site for Localization 2. My last lab ironically left Click somewhat deconstructed, so I spent a while getting the tower setup again and everything onboard and connected. I've been working on wall-following, an important and somewhat time-consuming first step for the actual implementation. I'm using just the sonar to follow the right wall, so most of the time it will be pointed in that direction. It's not quite working, but should only require maybe a half hour more of tweaking. I'm working on adapting my light-following code from the first lab to the task, as that did a pretty good job of making continuous, smooth directional updates while always moving forward. It should work with the sonar input. I'm still debating the use of odometry. I'm thinking maybe I can have one of two things happen to trigger the robot thinking it's entered a new area of the map. The first is if the right wall suddenly falls away, implying an open doorway or hall, or if its bump sensor is triggered. The second is some sort of regular update, where the robot stops and looks forward, left, and right to see if anything's changed and update probabilities. This could be triggered either by odometry or by a time-based mechanism. I'm actually leaning a bit towards the latter. If I can keep the robot moving at relatively the same speed, which experience has dictated it's fairly good at doing, this should be reliable enough. The timer gets reset whenever the robot stops for either reason. I'm going to give this approach a try first, as it should be quicker and easier to implement. I'll make the interval user-specifiable via the menu wheel.

4/16/04 - I've got wall following working fairly well. It's based on a couple of parameters which the user can set interactively when the robot first starts up. One is the distance it should try and maintain from the wall. The other is the threshold, or the amount of change it needs to see before it reacts to a variation in the input from the sonar. Right now it's probably too sensitive, see the video to watch it follow the wall a little too well, but by tweaking the values I should be able to get it to work pretty smoothly. I've got enough now that I should move on, anyway, and tweak more as necessary once I get the maze built and mapped.

Some media:
Video of Click following walls (5.55mb)
Picture of Click
Another picture


Perspective

Extinguisher
So far, it seems to be doing well. I am taking it into the lab for another round of tests and some picture taking this morning. With the thresholds properly set, it really does a nice job transitioning between modes and blowing out the candle. I enjoy how the robot slows to a crawl and employs its fan when it gets near the light. It is also rewarding to see it start back up and try again if it fails to completely blow out the candle on the first try. It could use a better light-following algorithm, more specifically one that doesn't put as much emphasis on the light, as it can easily be trapped by a malicious maze designer if the light comes over the top of a v-shaped obstacle. Its reflexes cause it to turn in order to try and get out, but then once it gets attracted by the light and heads back for it. As I have mentioned above, I feel like this was a good project in preparation for a more challenging navigation problem. Most likely I will attempt navigation with a map over the next three weeks, and then mapping for the remainder of the course.

Localization 1
I'm pleased with the progress I've been able to make in preparation for the final project. I spent a good deal of time researching the sonars, the IR rangefinders, and the handyboard in hopes of making them all work together, and learned quite a bit about each along the way. I'm also pleased with the changes I made based on this knowledge. I think my simpler, more straightforward design will pay off in the long run. I do wish I could get the IR sensor to work properly, as I'm interested in working with the wall-hugging robot. I think it will be a lot faster and more solid, as the sonar-only version will have to go for a bit, check left and right, then go for a bit more, and will probably be more prone to bumping into things and generally less graceful. Something about the assymetry also appeals to me.