The Potential Field Approach (PFA) [1] is a well known method in reactive robot navigation. The basic idea simply models (detected) obstacles as repulsors and goal(s) as attractor(s) depending on their proximity to the robot. The rotational of the resulting field returns the motion vectors for our mobile. PFA are intuitive, simple, smooth and, up to a point, reliable.

Despite being relatively old, this technique is still widely used in navigation. Indeed, many robotic wheelchairs use PFA to avoid collisions by removing control from the user when there is present danger in what we call "safeguard operation mode". Still, PFA per se can not be used for much more in its original formulation due to well reported problems: i) oscillations in corridor-like situations; ii) sensitivity to local minima; and iii) local traps. The following viewo shows the commented problems:

These problems can be solved by more powerful, enhanced versions of PFA, mostly based on making them less reactive. Algorithms used in assistive navigation instead of PFA include the Dynamic Window Approach and the Vector Field Histogram (VFH).


[1] Khatib, O., "Real-Time Obstacle Avoidance for Manipulators and Mobile Robots", Int. J. of Robotic Research, Vol.5, No.1 (1986), p.60

Autonomous navigation initially followed the so called sense-plan-act scheme (SPA) [Albus91], that work with a model of the environment. SPA has some well known drawbacks like strong dependence on a correct model of the environment and high latency. Reactive control architectures [Brooks86], instead, create simple behaviours by coupling sensor readings and actions and complex ones by combining several basic behaviours running concurrently. Reactive behaviours are fast, quite robust against sensor errors and noise and can easily adapt to changes in hardware or tasks. Yet, emergent behaviours are unpredictable, not necessarily efficient and prone to fall into local traps. Hybrid schemes like the well known 3T control architecture [Bonasso97] solve the aforementioned problems by combining both reactive and deliberative paradigms to achieve the best possible performance.

In order to measure how well a device is or a person deals with a given vehicle, many people work on predefined obstacle courses. Kilkens et al. reviewed a wide array of wheelchair skill tests, aimed to asses the ability to propel and maneuver a wheelchair under standardized
and/or simulated conditions of daily living. Skills included in the 24 tests in order of frequency were:

- Wheelchair propulsion, assessed in terms of period of time, a fixed distance or longest distance possible.
- Transfer from and to the wheelchair, usually examining the performance in different transfers.
- Negotiation of kerbs, with kerb's height ranging from 0.025 to 0.15m, and some requiring ascending and descending the kerb.
- Ascending slopes, defined in terms of inclination (ranging from 1 to 11 degrees) and length (ranging from 3 to 21m).
- Traversing tracks, e.g. slalom, figure of eight and obstacle course.
- Sprinting over a fixed distance (ranging from 6.5 to 30m).
- Performing a wheelie was also a skill included in few of the tests assessed.
- Half of the tests included other specific wheelchair skills, e.g. managing brakes, negotiating doors or loading the wheelchair into a car.

In eight of these tests, wheelchair skills were a part of a broader measure of ADL skills like eating, bed mobility skills and washing hands. Ideally, tests have to be as efficient and as short as possible, and should not require much space or special equipment. The VFM or the TAMP take like 1 hour to complete, while the WST needs only 30 min.

The most common outcome of these tests are values like task performance time, Physical strain, Independence in performance, distance covered in propulsion, endurance and other subjective ratings like perceived difficulty. Kilkens argues that tests should have preferably a simple scoring system, convenient to use and easi to analyze. It is important to note that these metrics can be applied mostly to obstacle courses but usually not to daily living, as they are measured over a given time and trajectory.



The main problem with most of these courses is that they require a large space and investment to build structures just according to specifications. Furthermore, they are used to measure skills in controlled situations, rather than in everyday ones. Still, they are widely used because they provide a benchmark to test most wheelchairs models.

I know, I know. It's a Russian thing.
When we're about to do something stupid,
we like to catalog the full extent
of our stupidity for future reference.
- S. Ivanova, A Voice in the Wilderness

And just when everything seems to be going so well, someone always asks, "so, what is so special about your work?". For scientists, a simple "it works" does not seem enough, so everyone stomps into the field of metrics sooner or later. Inventing some metrics of your own, while obviously appealing, will just not do. Since words might do you no good either, this is the point where standards come handy. The key idea here is that we are good as long as our system behaves better than previous, similar ones in at least a few in quantifiable aspects.



When one starts to look for standard standards in the wheelchair navigation field, bad news come first: there seems to be no established one to measure wheelchair performance -specially regarding power wheelchair navigation and, more specifically, shared control-. Fortunately, there is a large number of proposals that more or less agree regarding parameters of interest in assisted navigation (e.g. [webster et al, 88]).

Assisted wheelchair navigation is a field where many fields converge, from cognitive sciences to medicine, and all the way through engineering. Consequently, trying to fit all related metrics in the same bin would be like trying to explain feelings with differential equations. Instead of doing so, we will go for a tentative distinction between different categories, which might be more or less correlated, but are simpler to explain separately. Keep in this channel for the categories very soon in a web near you :)

The easiest solution to the localization problem is to use dead reckoning: if we know how fast and how long we have been moving, we can estimate approximately how far we are from our original position. Alternatively, we can measure the number and length of our steps or the number of wheel turns of our vehicle to do the same. Indeed, boats used to proceed like that: their speed and direction was estimated in terms of how long a knot in a rope dropped from the bow of the vessel took to reach its back side (1) and new locations were estimated in terms of how long that speed was kept. Naturally, odometrics can only work for a time, because errors tend to accumulate in an unbounded way. If no external reference is used to correct the position from time to time, after a while the vehicle is lost. One could try, for example, to precalculate the path to a spot 10 m away and then try to walk there blindfolded. Chances are one would most likely end fairly far from the desired destination, specially if turns are involved, or, at worst, stuck into an unexpected obstacle. In order to avoid this problem, we could just open the eyes that, in the robot case, means we can combine different information sources using methods like Kalman filters or Montecarlo techniques to remove or, at least, reduce uncertainty.

Ok, so maybe dead reckoning was not such a good idea...

Things are easier if we know the location of at least two relevant features in the environment that we can perceive at a given location. For example, boats originally relied on the position of the stars and nowadays they use GPS satellites to triangulate where they are. Mobile robots can do the same if external beacons are available (GPS, ultrasounds, wireless beacons, etc).

Of course, active beacons might not be available; GPS, for example, can not be used indoors. If this is the case, robots can refer to known features in the environment -natural or artificial landmarks- to estimate where they are, in the same way we could check a metro station we are watching in a map to decide where we are at the moment. In these cases, cameras are typically used not as range sensors, but to detect the expected landmarks within the field of view e.g. (Urdiales et al, 2009).
A very special case among these ones is omnicameras, whose optics have been designed so that all points in the world are projected through a single center of projection. Hence, they capture 360º fields of view in a single frame and usually extract fairly reliable landmarks from the picture (see Keith Price bibliography on the subject).

If cameras are not available or do not want to be used, geometric landmarks can also be detected. These landmarks are significant, distinct locations that can be detected by means of range sensors, like a corridor with two open doors at each side. This would be equivalent to try to check where we are by touching the walls if the lights go out. However, similar landmarks may exist in different areas of the environment, so it is necessary to disambiguate perceptions by keeping some track of the mobile position on accumulating several landmarks in a row via a statistic method like, say, Markov Model e.g. (Baltzakis, 2003)(Fox, 1998).

In absence of a model of the environment, it is necessary to correct the position, store detected landmarks and build a new model, all at the same time. Think, for example, of trying to draw a map of a city we have never been in by marking every distinctive building we see and trying to guess the distance and relative position between each two of them. This problem is known as SLAM. SLAM conforms a quite complex, complete field of research and it is out of the scope of our work.

---

(1) In fact, the idea was to sing a song and write down in the bitacora in which word it had stopped. One can only hope that the crew was able to keep the rythm.

Traditionally, wheelchairs relied on range sensors to navigate. These sensors -in decreasing order of cost, weight and range- include laser, sonar, infrared and bumpers, although cameras have been also used to measure distance either in pairs (stereovision/divergence) or via optical flow.

TOF sensor fusion

Most range sensors basically use the TOF (Time of Flight) of an echo signal to offer their distance to the closest obstacle in the direction of the sensor, but many of these can be rotated like a sonar in a submarine to cover a wider range of detection. Others, like cameras, provide information on a wider area, but this information is typically more complex to process. In a dynamic, potentially unstructured environment, a quick response time might be the difference between safety and collision, so in many cases, when video cameras are used, they are combined with other range sensors to achieve faster responses. Furthermore, visual information is so rich that video processing has only been solved when some constraints can be applied. These restrictions usually imply some knowledge about the operation environment and a heavy specific problem-solving orientation. Simpler range sensors also present their own drawbacks. Sonar sensors, for example, have an uncertainty angle that, in models like Polaroid, may be up to 22.5º. Consequently, when obstacles are at a significant distance, say 4-5 m, we know how far they are, but not exactly where. Infrarred sensors are sensitive to natural light and their output may change sharply when there is a significant illumination change, like whenever someone is stepping in front of a window. switches on a light or opens a door. Lasers can not detect glass doors and deal poorly with black surfaces. Furthermore, all three of these sensors only detect obstacles in their own plane, so irregular objects like a table might be completely missed. Recently, though, laser sensors have become smaller, cheaper and stronger, so they have been particularly favored by the wheelchair industry. A classic SICK LMS 291-S05 laser weights 4.5kg and is 156 x 155 x 210mm, whereas a modern Hokuyo URG-04LX weights 160g and is just 50 x 50 x 70mm.

Hokuyo laser

In any case, since all sensors have advantages and drawbacks, it is usual to work with several ones or, at least, to statistically combine the readings of a single one in time and space to obtain more reliable knowledge on the environment. If these readings are combined in a short time span within the surroundings of the robot, range sensor readings can be used to avoid close obstacles by heading it in a free direction. Readings can also be combined into a wider model of the environment to predict more efficient and safer trajectories to the goal, to combine several goals or to coordinate different mobiles. In order to build global models of the environment, though, it is important to know the position of the mobile within the environment with some accuracy. Otherwise, it would be like asking directions to someone who is actually lost.