Designing Autonomous Mobile Robots phần 6 ppsx

164 Chapter 11 The correction we will actually make for each axis is calculated as it was for azimuth, by multiplying the full implied correction (error) by the Q factor for the whole fit. x COR = – (Q FITx * E x ) (Equation 11.13) And y COR = – (Q FITy * E y ) (Equation 11.14) Assume that this robot had begun this whole process with significant uncertainty for both azimuth and position, and had suddenly started to get good images of these col- umns. It would not have corrected all of its axes at the same rate. At first, the most improvement would be in the azimuth (the most important degree of freedom), then in the lateral position, and then, somewhat more slowly, the longitudinal position. This would be because the corrections in the azimuth would reduce the azimuth uncertainty, which would increase the observation quality for the position measurements. However, we have only discussed how uncertainty grows, not how it is reduced. Other profiles The equations discussed thus far are the most primitive possible form of fuzzy logic. In fact, this is not so much a trapezoidal profile as a triangular one. Even so, with the correct constants they will produce very respectable results. Figure 11.12. Triangular vs. trapezoidal profile Error Q factor Error = 0 Error = Uncertainty 1.0 Triangular Profile Q = (U-E)/U 0.0 E1 E2 Q1 Trapezoidal Profile 165 Hard Navigation vs. Fuzzy Navigation Figure 11.12 shows another possibility for this function. In this case, we simply refuse to grant a quality factor greater than Q1, so the navigation will never take the whole implied correction. Furthermore, any error that is greater than E2 is considered so suspect that it is not believed at all. If the breakpoints of this profile are stored as blackboard variables, then the system can be trimmed during operation to product the best results. The referenced state It should be obvious that the navigation scheme we are pursuing is one that requires the robot to know roughly where it is before it can begin looking for and processing navigation features. Very few sensor systems, other than GPS, are global in nature, and very few features are so unique as to absolutely identify the robot’s position. When a robot is first turned on, it is unreferenced by definition. That is, it does not know its position. To be referenced requires the following: 1. The robot’s position estimate must be close enough to reality that the sensors can image and process the available navigation features. 2. The robot’s uncertainty estimate must be greater than its true position error. If our robot is turned on and receives globally unique position data from the GPS or another sensor system, then it may need to move slightly to determine its heading from a second position fix. Once it has accomplished this, it can automatically become referenced. If, however, a robot does not have global data available, then it must be told where it is. The proper way to do this is to tell the robot what its coordinates and azimuth are, and to also give it landmarks to check. The following incident will demonstrate the importance of this requirement. Flashback… One morning I received an “incident report” in my email with a disturbing photo of a prone robot attached. A guard had stretched out on the floor beside the robot in mock sleep. Apparently, the robot had attempted to drive under a high shelving unit and had struck its head (sensor pod) on the shelf and had fallen over. This was a very serious—and thankfully extremely rare—incident. Those who follow the popular TV cartoon series “The Simpsons” may remember the episode in which Homer Simpson passed down to his son Bart the three lines that had 166 Chapter 11 served him well his whole life. The third of these lines was, “It was that way when I found it, boss.” In the security business, an incident report is an account of a serious event that contains adequate detail to totally exonerate the writer. In short, it always says, “It was that way when I found it, boss.” This incident report stated that the robot had been sent to its charger and had mysteri- ously driven under the shelving unit with the aforementioned results. In those days we had a command line interface that the guard at the console used to control the robot. A quick download of the log file showed what had happened. Instead of entering the command, “Send Charger,” the operator had entered the command “Reference Charger.” This command told the robot its position was in front of the charger, not in a dangerous aisle with overhead shelves. The reference program specified two very close reflectors for navigation. The robot was expected to fine tune its position from these reflectors, turn off its collision avoidance, and move forward to impale itself on the charging prong. Since there was no provision in the program for what to do if it did not see the reflectors, the robot simply continued with the docking maneuver. The oversight was quickly corrected, and later interlocks were added to prevent the rep- etition of such an event, but the damage had been done. Although this accident resulted in no damage to the robot or warehouse, it had a long-lasting impact on the confidence of the customer’s management in the program. If the robot’s uncertainty estimate is smaller than the true error in position, then the navigation agents will reject any position corrections. If this happens, the robot will continue to become more severely out of position as it moves. At some uncertainty threshold, the robot must be set to an unreferenced state automatically. If the robot successfully finds the referencing features, then—and only then—can it be considered referenced and ready for operation. Certain safety related problems such as servo stalls should cause a robot to become unreferenced. This is a way of halting automatic operations and asking the operator to assure it is safe to continue operation. We found that operators ranged from extremely competent and conscien- tious to untrainable. Reducing uncertainty As our robot cavorts about its environment, its odometry is continually increasing its uncertainty. Obviously, it should become less uncertain about an axis when it has received a correction for that axis. But how much less uncertain should it become? 167 Hard Navigation vs. Fuzzy Navigation With a little thought, we realize that we can’t reduce an axis uncertainty to zero. As the uncertainty becomes very small, only the tiniest implied corrections will yield a nonzero quality factor (Equation 11.9 and 11.10). In reality, the robot’s uncertainty is never zero, and the uncertainty estimate should reflect this fact. If a zero uncertainty were to be entered into the quality equation, then the denominator of the equation would be zero and a divide-by-zero error would result. For these reasons we should establish a blackboard variable that specifies the minimum uncertainty level for each axis (separately). By placing these variables in a blackboard, we can tinker with them as we tune the system. At least as importantly, we can change the factors on the fly. There are several reasons we might want to do this. The environment itself sometimes has an uncertainty; take for instance a cube farm. The walls are not as precise as permanent building walls, and restless denizens of the farm may push at these confines, causing them to move slightly from day to day. When navigating from cubical walls, we may want to increase the minimum azimuth and lateral uncertainty limits to reflect this fact. How much we reduce the uncertainty as the result of a correction is part of the art of fuzzy navigation. Some of the rules for decreasing uncertainty are: 1. After a correction, uncertainty must not be reduced below the magnitude of the correction. If the correction was a mistake, the next cycle must be capable of correcting it (at least partially). 2. After a correction, uncertainty must never be reduced below the value of the untaken implied correction (the full implied correction minus the correction taken). This is the amount of error calculated to be remaining. Uncertainty may be manipulated in other ways, and these will be discussed in upcoming chapters. It should be apparent that we have selected the simple example of column navigation in order to present the principles involved. In reality, features may be treated discretely, or they may be extracted from onboard maps by the robot. In any event, the principles of fuzzy navigation remain the same. Finally, uncertainty may be manipulated and used in other ways, and these will be discussed in upcoming chapters. Sensors, Navigation Agents and Arbitration 12 CHAPTER 169 Different environments offer different types of navigation features, and it is important that a robot be able to adapt to these differences on the fly. There is no reason why a robot cannot switch seamlessly from satellite navigation to lidar navigation, or to sonar navigation, or even use multiple methods simultaneously. Even a single- sensor system may be used in multiple ways to look for different types of features at the same time. Sensor types Various sensor systems offer their own unique strengths and weaknesses. There is no “one size fits all” sensor system. It will therefore be useful to take a very brief look at some of the most prevalent technologies. Selecting the proper sensor systems for a robot, positioning and orienting them appropriately, and processing the data they return are all elements of the art of autonomous robot design. Physical paths Physical paths are the oldest of these technologies. Many methods have been used to provide physical path following navigation. Although this method is not usually thought of as true navigation, some attempts have been made to integrate physical path following as a true navigation technique. The concept was to provide a physical path that the robot could “ride” through a feature poor area. If the robot knew the geometry of the path, it could correct its odometry. Commercial AGVs, automatic guided vehicles, have used almost every conceivable form of contact and noncontact path, including rails, grooves in the floor, buried wires, visible stripes, invisible chemical stripes, and even rows of magnets. 170 Chapter 12 The advantages to physical paths are their technical simplicity and their relative ro- bustness, especially in feature-challenged environments. The sensor system itself is usually quite inexpensive, but installing and maintaining the path often is not. Other disadvantages to physical paths include the lack of flexibility and in some cases cos- metic issues associated with the path. A less obvious disadvantage for physical paths is one that would appear to be a strength: their precise repeatability. This characteristic is particularly obvious in upscale environments like offices, where the vehicle’s wheels quickly wear a pattern in the floor. Truly autonomous vehicles will tend to leave a wider and less obvious wear footprint, and can even be programmed to intentionally vary their lateral tracking. Sonar Most larger indoor robots have some kind of sonar system as part of their collision avoidance capability. If this is the case, then any navigational information such a system can provide is a freebee. Most office environments can be navigated exclu- sively by sonar, but the short range of sonar sensors and their tendency to be specularly reflected means that they can only provide reliable navigation data over limited angles and distances. There are two prevalent sonar technologies: Piezo-electric and electrostatic. Piezo- electric transducers are based on crystal compounds that flex when a current passes through them and which in turn produce a current when flexed. Electrostatic transducers are capacitive in nature. Both types require a significant AC drive signal (from 50 to 350 volts). Electrostatic transducers also require a high-voltage bias during de- tection (several hundred volts DC). This bias is often generated by rectifying the drive signal. Both of these technologies come in different beam patterns, but the widest beam widths are available only with Piezo-electric transducers. Starting in the 1980s the Polaroid Corporation made available the narrow beam Piezo-electric transducer and ranging module that they had developed for the company’s cameras. The cost of this combi- nation was very modest, and it quickly gained popularity among robot designers. The original Polaroid transducer had a beam pattern that was roughly 15 degrees wide and conical in cross-section. As a result of the low cost and easy availability, a great number of researchers adopted this transducer for their navigation projects. To many of these researchers, the configuration in which such sensors should be deployed seemed obvious. Since each transducer covered 15 degrees, a total of 24 transducers 171 Sensors, Navigation Agents and Arbitration would provide the ability to map the entire environment 360 degrees around the robot. The obvious way is usually a trap! Sonar Ring 15 o The Polaroid “ring” configuration ignored the principles of proper sensor deploy ment. Figure 12.1. The ubiquitous “Polaroid ring” sonar configuration 172 Chapter 12 The most obvious weakness of the narrow-beam ring configuration is that it offers little useful information for collision avoidance as shown in Figure 12.2. This means that the cost of the ring must be justified entirely by the navigation data it returns. From a navigation standpoint, the narrow beam ring configuration represents a classic case of static thinking. The reasoning was that the robot could simply remain sta- tionary, activate the ring, and instantly receive a map of the environment around it. Unfortunately, this is far from the case. Sonar, in fact, only returns valid ranges in two ways: normal reflection and retroreflection. Figure 12.2. Vertical coverage of a narrow-beam ring 90 o Convex Object 90 o Wall (flat vertical surface) Normal reflection Retro-reflection Normal reflection Sonar Ring 1 6 5 4 3 2 Phantom Object Table Steps Figure 12.3 Modes of sonar reflection 173 Sensors, Navigation Agents and Arbitration Figure 12.3 shows two cases of normal reflection, one from a flat surface and one from a convex object. Also shown is a case of retroreflection from a corner. Retroreflec tion can occur for a simple corner (two surfaces) or a corner-cube. In the case of a simple corner such as in Figure 12.3, the beam will only be returned if the incidence angle in the vertical plane is approximately 90 degrees. Unless the wall surface contains irregularities larger than one-half the wavelength of the sonar frequency (typically 1 to 3 cm), sensors 2, 3, and 5 will “go specular” and not return ranges from it. Interestingly, beam 2 returns a phantom image of the post due to multiple reflections. Sonar Ring Wall Wall Corner Post Phantom of post Figure 12.4. Sonar “map” as returned by narrow-beam sonar ring Figure 12.4 approximates the sonar data returned from our hypothetical environ ment in Figure 12.3. If the data returned by lidar (Figure 10.2) seemed sparse, then the data from sonar would appear to be practically nonexistent. [...]... the cardinal directions of the robot have been used very successfully by academic researchers and commercial manufacturers of indoor mobile robots Polaroid has also introduced other transducers and kits that have been of great value to robotics developers For Cybermotion robots, on the other hand, we selected wide-beam Piezo-electric transducers Our reasoning in this selection was that the resolution... z, but it may not be possible to have any two of these at the same time Since Polaroid transducers have been the most popular sonar sensors for mobile robots, it is instructive to look at some of the configurations into which they have been deployed Figure 12 .6 shows a configuration that resulted from the recognition of the importance of focusing on the forward-direction This configuration has sacrificed... Pain, Fear and Confidence In earlier chapters, we have seen parallels between the ways animals, people, and robots cope with problems such as action planning and navigation At this point, it is useful to consider the purpose of certain basic emotions in survival To a large extent, autonomous robots will be subject to the stresses analogous to those experienced by living creatures in that they both... wellplaced blow from her purse and a promise to report the offensive beast to show management1 There is a lesson here Bad things will happen to good robots and robots are presumed guilty unless proven innocent If there is ever an accident involving humans and robots, the robot is going to be blamed unless its keepers can provide absolute documented proof that the incident was caused completely by the human... signal-processing process Avoid fixed thresholds and filter parameters 6 Never discard useful information For example, if the data from a sonar transducer is turned into a range reading in hardware, it will be difficult or impossible to apply additional data filters or adaptive behaviors later 7 Consider how sensor systems will interact with those of other robots 8 Sensor data processing should include fail-safe... knowledge, and that the wider beam width allowed us to image a larger volume of space per ping The signal of a sonar system is reduced with the square of the pattern radius, so the 60 -degree beam pattern we selected returned 1/16th the energy of the 15-degree transducers To overcome the lower signal-to-noise rations resulting from wide-beam widths, we used extensive digital signal processing A final warning... perpendicularity to the beams In the patterns of Figure 12 .6, there are 12 opportunities for a wall to be within the range of the transducers and not be seen For this reason, when two halfrings are used, the lower set of transducers is usually staggered in azimuth by one-half the beam-width angle of the devices Notice that this configuration has 26 transducers to fire and read If the range of the transducers... per a half-ring from 13 to 16 (as in Figure 12.7) or even to 25, depending on the overlap desired By this point, we can see that getting good collision avoidance coverage with a ring configuration is a bit like trying to nail jelly to a tree Everything we do makes the problem worse somewhere else Figure 12.7 Staggered-beam configuration 181 Chapter 12 Several commercial robots have achieved reasonable... Cybermotion, Inc.) Lidar has occasionally been used in a downward mode to illuminate an arc in front of the robot Configurations like that shown in Figure 12.11 have been used by several robots, and have even been added to Cybermotion robots by third parties Side View Lidar Obstacle Top View Figure 12.11 Downward-looking lidar for collision avoidance 1 The upper trace is the raw sonar echo data and the lower... transducers can bounce around the natural corner reflectors formed by the walls, floor, and ceiling and then return to the robot long after they were expected Additionally, pings from other robots can arrive any time these robots are nearby The signal processing of the sonar system must be capable of identifying and ignoring these signals If this is not done, the robot will become extremely paranoid, even . commercial manufacturers of indoor mobile robots. Polaroid has also introduced other transducers and kits that have been of great value to robotics developers. For Cybermotion robots, on the other hand,. floor. Truly autonomous vehicles will tend to leave a wider and less obvious wear footprint, and can even be programmed to intentionally vary their lateral tracking. Sonar Most larger indoor robots. remember the episode in which Homer Simpson passed down to his son Bart the three lines that had 166 Chapter 11 served him well his whole life. The third of these lines was, “It was that way when