Tuesday, 11 January 2011

There's No Prospective Information About Friction, or, Why I Fell Over on the Ice

In which I justify why I, a healthy perceiver-actor, slipped and fell on a clearly visible icy patch, breaking my wrist for the second time, using SCIENCE.

It's been a cold, icy winter here this year, and 6 weeks ago I slipped on a patch of ice and fell entirely on my (previously broken) wrist. The ensuing physics did enough damage that I needed surgery to set the wrist with two pins, and I am only today out of the cast. These kinds of falls and injuries are very common; half of all falls  in the US are caused by insufficient friction, and the types of injuries (broken wrists and collarbones, etc) suggest reactive responses to the slip - people using their arms to try and regain a sudden, unexpected loss of balance. 

The two papers I'm going to talk about are from the lab of my favourite developmental psychologist, Karen Adolph, who has done some excellent affordance work using the transition from crawling to walking as a way to studying the changing perception-action performance of children. This research, however, asked about whether a perceiver can detect information about upcoming friction conditions and use this information for prospective control. The answer seems to be no, because there isn't any information. Given that action requires information, the absence of information might explain the often catastrophic failures of action we see on ice and other low-friction surfaces.

At least, that will be my story.

First, we need to define some terms.

Prospective control is about shaping our behaviour to cope with upcoming circumstances (for instance, adjusting your stride length to hit a set of stairs just so). Our bodies have significant mass and hence inertia, making it impossible to 'turn on a dime'; all adjustments take time and thus must be planned out ahead of time.  In addition, the transition from one action to another is potentially perturbing to your stability - that change in stride length may make you likely to trip. It's important that these transitions are managed carefully so they occur smoothly and under full control.

Planning a change to an action requires information that a change is required; no information, no change. (This is what makes prospective control a different idea from prediction, which is the more cognitivist hypothesis that planning for the future is based not on information, but on knowledge.) Birds occasionally fly headlong into clear glass windows without making any attempt to slow down because they simply don't see the window and the frame just specifies a passable gap. Knowledge of windows also doesn't help: I've walked into clear glass sliding doors twice in my life, and I "know" how windows work. 

So the perception-action system as a whole exhibits inertia - it will tend to continue doing what it's currently doing until pulled into doing something else. This pull has to be 'strong' enough to overcome the system's inertia, or, to put it another way, the new affordance must be clearly specified and this information detected in good time (and even this is not necessarily enough, when the cost of persisting is low enough; Kent et al, 2009). The prospective control of locomotion therefore requires that there is information about upcoming surface properties, specifically its affordances for locomotion.

What about the property friction? Friction is the resistive force that emerges as two surfaces rub against one another. Friction is therefore not a property of a single surface; it is a property of a given surface-surface combination, and thus is not uniquely defined until the surfaces are in contact. Worse (from the point of view of prospective control), the specific co-efficient of friction (CoF) that emerges depends on a staggeringly wide array of factors, including but not limited to, the temperature and velocities of the surfaces, atmospheric conditions, as well as various geometric properties (the size and shape of the objects, the slope of the surfaces), not to mention which two surfaces are in contact. In general, the CoF cannot be computed but must be measured empirically, and only once the surfaces are in contact.

This analysis of the physics of the situation suggests that friction cannot be specified ahead of time, because it is not uniquely defined ahead of time. This suggests that there can be no information about friction, and hence no possibility of prospective control with regard to it. This would certainly explain why we can fall so easily on low-friction surfaces, even though there is clear motivation to learn not to. It doesn't explain why we don't always fall; with care, you can walk on a low friction surface. However, I think the key term there is with care; by slowing down, placing your feet carefully and lowering your centre of mass you can reduce the consequences of any given slip and thus give yourself time to respond reactively to conditions.

This only works if the consequences allow you the required time, which is never guaranteed. Adolph's lab was therefore interested in the question of prospective control, so they investigated whether there was anything available to perception, even a mere cue, that people might come to rely on given the absence of actual information.

Prospective visual perception of slipperiness
Joh et al (2006) ran four studies. The first simply asked people 'By only looking at it, how would you tell if the floor was slippery?'. 91% of participants responded with something referring to the shine of a surface, and 58% mentioned a contaminant of some kind (e.g. water or grease). This clear awareness of shine seems sensible; surfaces likely to be slippery are often smooth or contaminated with something that will reflect light well. Of course, shine is not lawfully connected at any scale to slip (there are numerous counter examples of surfaces which are dull but slippery, e.g. sand on a floor, as well as surfaces which are shiny but not slippery, e.g. linoleum floors in office buildings). This means (by the Turvey, Shaw, Reed & Mace (1981) approach) that shine cannot serve as information for slip. Worse, shine perception is highly variable, and shows little constancy over common variations in viewing conditions (e.g. ambient lighting).

The second study therefore asked, do people seem to be actually using shine to judge slipperiness? Two groups made judgements from three distances of either shine or slip of surfaces which varied in colour and gloss. The static CoF increased with gloss but remained constant with colour and viewing distance; however, judgments of both shine and slip varied with all these factors, and in a similar fashion. Shine and slip seem to be judged using the same, unreliable cue. This effect was replicated under outdoor lighting conditions in the third study. The final study asked people to choose which of these surfaces they would prefer to walk on (judging the affordance); judgements were driven almost entirely by shine, the perception of which continued to be affected by the colour and distance of the surface (factors irrelevant to friction).

People seem to be using shine to judge slip, but a) shine perception was unreliable and influenced by factors which don't impact on friction, and b) it was only partly related to slip anyway. These are only judgement studies, but they reveal that people are being forced to rely on a mere visual cue (shine) for evaluating upcoming frictional conditions, and suffering because shine perception is highly variable and shine doesn't specify friction anyway.

Prospective non-visual perception of slipperiness
Most of the prospective control in locomotion is visual control, because vision provides access to information (if any) about surface conditions far enough in advance to be of use for smooth online control of action. However, every step you take provides non-visual information about the surface you are walking on, and importantly it potentially provides information about friction, because there actually are two surfaces in contact with each other. Can this information be used to control future steps, i.e. for prospective control?

Joh et al (2007) had adults stand on flat low, medium and high friction surfaces, and make judgements about whether they could walk down the same surface once it sloped. The experiments compared judgements of which slopes the people thought they could traverse to the static CoF between that surface and a lead block control. Experiment 1 had people simply stand on the flat surface; Experiment 2 instructed them to explore the surface by sliding their feet around while standing in place; Experiment 3 hid the slope and varied friction with different shoes, and also asked people to judge the steepest slope they could stand on and walk up. Experiment 4 tested whether the lead block was an adequate control by having people stand on various slopes after having made judgements from the flat surface.

The results were broadly consistent; participants were extremely poor at identifying when they could safely traverse the sloped surface, and the exploration did not particularly help. Not only were judgements highly variable, but worse, the errors were in the least useful direction. Participants systematically underestimated their ability to walk down high friction slopes, and overestimated their ability to walk down low friction surfaces. I'm inclined to think this reflects the type of conservative, regression to a mean type behaviour that is typical under conditions of such extreme uncertainty; given that there seems to be no useful information for friction, people simply pick a value and work to that.

The only condition that improved judgements was Experiment 5, in which people made judgements about traversing the slope with one foot on the flat surface and one actually on the slope; people were better here, although they still made errors with the high friction surfaces.

In general, there appears to be no information available about friction for the prospective control of locomotion. Even when in contact with the flat surface and actually generating frictional forces, people were unable to use this to identify which slopes they could then traverse. A big part of this reason is that friction is simply so complicated; the actual CoF for any given situation depends on too many variables (as described above). In the specific context of locomotion, factors such as how the foot is placed and wear on the shoe can have effects, and the net result is that any information generated by one step is simply not informative about the next one.

Perceptual learning requires a fairly stable target: invariants-over-transformation as proposed by Gibson are thus ideal candidates, but when the physics of the situation means such invariants do not exist, perceptual learning simply cannot happen. This is a great example of how information trumps physics, as far as a perceiver is concerned; friction exists but does not lawfully give rise to any invariants in any energy arrays which might specify that friction, and thus the perceiver has no access to the physical property or related affordances.

So why did I fall over?
I saw the ice coming, but that never entered into the control of my locomotion and I hit it with a stride that had been appropriate for the high-friction dry pavement but that was entirely inappropriate for ice, especially in my rigid hiking shoes which slip on ice a lot. There was no information that enabled me to alter the control of my steps, and the cue of shine and the knowledge that there was ice did not arrive in time for me to make any strategic changes to my locomotion (e.g moving to avoid the icy patch). So I fell because changes in action require information, and in the absence of information it was impossible for me to make that change. Shine isn't information, and neither is the knowledge that there was ice, because neither is supported by the kind of lawful relation that perception requires.

Joh AS, Adolph KE, Campbell MR, & Eppler MA (2006). Why walkers slip: shine is not a reliable cue for slippery ground. Perception & Psychophysics, 68 (3), 339-52 PMID: 16900828 Download

Joh AS, Adolph KE, Narayanan PJ, & Dietz VA (2007). Gauging possibilities for action based on friction underfoot. Journal of Experimental Psychology. Human Perception and Performance, 33 (5), 1145-57 PMID: 17924813 Download

Kent, SW, Wilson, AD, Plumb, MS, Williams, JH, Mon-Williams, M (2009) Immediate movement history influences reach-to-grasp action selection in children and adults. Journal of Motor Behavior, 41, 10-15. Download


  1. While I do agree that friction is a physical property that cannot be directly perceived with any accuracy, this is sort of why these Gibsonian ideas fall apart for me. No single visual provides a good indication of an object's friction (just as no visual cue, other than relative size, provides any direct indication of an item's veridical weight). But still we cope (pretty successfully) with friction in the context of prehension.

    As I currently live in Southern Ontario, where the ground is covered in snow and ice for over 6 months of the year, perhaps I can suggest an alternative (but almost entirely anecdotal) hypothesis. Slipping to me is a rather binary thing - either you successfully traverse the ground you are on, or you fail to traverse in catastrophic fashion. Errors you make (presumably errors in the angle of contact your foot makes to the ground/speed of gait) either are not made (since you don't slip), or are so bad that you have other things on your mind than consolidating and correcting a motor error (like how much my newly broken wrist hurts/how many pretty girls saw me fall over). Even so, you probably do better next time you are out. Living in the UK (and worse, NZ before that), where snow if comparatively infrequent, perhaps you just need more time in the snow to get fully tuned into the 'white stuff on ground could be slippery' framework. Of course, this will be a balancing act (oh ho), if you want to go anywhere in a decent time and maintain a degree to speed. But you will be giving a greater weighting to the possibility of falling with every passing occurrence (or near failure) that you encounter. These recent catastrophic failures may merely be a sad consequence of your lifetime of successful walking away from the snow and ice (glass doors beware). Adjust your prior!

  2. A couple of things:

    No single visual provides a good indication of an object's friction (just as no visual cue, other than relative size, provides any direct indication of an item's veridical weight).
    Actually size doesn't specify weight, and the single most robust 'illusion' in the literature is the size-weight illusion. Nothing visually specifies an object's weight unless it's moving, and then there's information everywhere - patch light displays of people lifting things produce very stable judgments of liftability, etc.

    But still we cope (pretty successfully) with friction in the context of prehension.
    This is true! But Joh et al (2006) address this, and note that the forces involved are low, and our fingers are able to move quickly to produce reactive corrections for prospective errors. Locomotion is slow and the forces are greater, so the consequences are greater and the odds of success are much lower because of less time, etc.

    Slipping to me is a rather binary thing - either you successfully traverse the ground you are on, or you fail to traverse in catastrophic fashion.
    This is a symptom of the problem. If there was information you could actually use to control your locomotion, you would have all the fine-grained control you normally have over walking behaviour. Instead, you step on ice and you either get lucky and step on it correctly, or you get unlucky, and fail. The control of action requires access to information; no information, no control, just luck, and that's not a lot of use.

    The learning aspect is indeed interesting. I'd be interested to see accident stats from various places that get ice reliably or not (although there's confounds like Yaktrax, etc).

    But whatever learning occurs, my bet (based on Adolph's analysis of the information) is that it won't actually produce improved real-time online locomotory control on ice. I think you'll see better stepping, slower walking, etc; these are more 'strategic' shifts, in the types of locomotion you choose, rather than 'control' shifts, in the way you produce that type. Does that distinction make sense?

    Adjust your prior!
    I did, and now I don't think Bayesian statistics are how we control action :)