As I was perusing Twitter, I bumped into this sponsored tweet by Shell, promoting Sensabot, a ruggedised remote ops robot designed for operations in dangerous environments by Carnegie Mellon’s NREC and now apparently adapted/adopted by Shell:
— Shell (@Shell) September 27, 2016
That sounds great… unfortunately, it strikes me, it misunderstands a couple of things – namely,
- fear, and the function it has in the human mind (not just the psyche – fear is a primarily neural response, secondarily perhaps cognitive and far behind it is any psychological aspect thereof), and
- what a robot ought to do/know.
Now, this might just be a marketing puff, though it probably is in its own way true – I have yet to see robots with a specific system catering for fear. It is also hardly just Shell and NREC I’m singling out here – the points are much more generic and have more to do with that robots do and what they therefore ought to understand about being human.
The gift of fear
When I was a child, I once managed to piss off my grandfather enough to have him lock me in the shed. The shed smelled of chicken feed, something that turns my guts upside down to this day. Now, the shed was pretty ok, all things considered, albeit dark and damp. However, I shared it with what at my youthful estimation must have been several hundred of small insects, spiders, centipedes and other creepy-crawlies.
I love nature. I just hate the things it sometimes produces. Safe to say if it does not have a brain and does creepy-crawly stuff, I will in all likelihood be creeped out by it.
So there I was, locked in with what I now know were more along the lines of a few thousands of these critters. I screamed like a banshee, until my grandmother took pity on me and released me.
Years and years later, memories of this crawled their way back into my mind at the most inopportune times. It was an uncharacteristically cool midsummer day, in the ageless and timeless beauty that you only get at NSC Bisley, the capital of UK target rifle shooting. Looking through the high-resolution scope of my rifle at Stickledown, the long distance (1,000+ yd) match rifle range, I was immersed in doing calculations of wind and gust patterns and adjustments and projectile drop in my head, trying to look out for any sign of crosswind at what is rightly known as one of the most treacherous ranges in the world of long distance target rifle shooting. That was when I suddenly felt that familiar crawl of a spider up my right leg. Had it not been for the fact that freaking out like a lunatic at a firing range while holding a stupidly overpowered sniper rifle in the middle of a few dozen other shooters on edge at what is to many of them a career match is generally shunned upon, I would probably have screamed. Instead, I unloaded, kicked the spider off me and tried to settle down, but by that time, it was all lost. My heart, pickled in adrenalin, pounded and pounded and I got some stupidly bad shots away before I conceded. And that’s how great my first Imperial Meeting went, back in 2006.
I was so bothered by this episode, I applied my usual method to it – reading every single book ever written on the subject of fear and anxiety responses (years later, when I would once again have to face scary memories from my past, much of that reading would prove helpful in retaining a modicum of sanity). I read tomes of evolutionary biology, a field so unfamiliar to me I had to ask a fellow student to give me the Cliff’s Notes of it. I read a fantastic book, The Gift of Fear by Gavin de Becker, in which he explains the importance of fear signals in avoiding violence. I read copiously on the physiology of fear responses, of the need for the inotropy and chronotropy of adrenaline, and the reason why battle cries and battle shouts are a universal feature of human civilisations.
And eventually, I came to realise that while fear probably lost me any stab I had at the Halford, the much coveted 1,100yd/1,200yd trophy that year (if we’re being honest here, any such chances were… fairly slim at best), it was a crucial part of getting my species, and quite probably the individual self, to that point. Of course, because this is not a Hollywood blockbuster about facing one’s fears and winning, I never went back to Bisley after this event and I would in fact never again shoot in a public competition. I would, however, spend plenty of time on the more fortunate end of a rifle, and never have another problem with it – even in inhospitable climates with various nasty creepy-crawlies.
Now, there’s a reason I’m giving a little personal vignette here – and that is to understand that we’re more familiar with the adverse effects of fear than we are with its ‘gift’, to use de Becker’s term. We, as a society, are in the mindset I was as I walked down the firing point and tried to figure out how the hell I am going to explain all this to my coach without becoming the club joke. Fear is bad. Fear is so bad, if you get too much of it, many opt for taking medication or seeking professional help (and that’s perfectly right so!). “Freedom from fear”, one of FDR’s often-quoted four freedoms, has put fear on the level of starvation, religious persecution and oppression of free expression. Fear is a big deal.
And justly so. Fear is, well, not nice. It puts people ‘on edge’, which is just fine, but is a prioritisation mechanism – it puts efficiency and survival ahead of communication and courtesy, and leads you to be perceived as unpleasant. Fear, especially long-term levels of heightened stress response (known sometimes the ‘biological embedding of anxiety’) can have utterly deleterious effects on long-term health and there seems now ample evidence that the damage is on a genetic level, i.e. capable of being passed on. The harm of fear thus becomes intergenerational.
At the same time, fear is necessary for humans, and it does not take much to think of a scenario when your entire ancestral line could have been wiped out if it had not been for your slightly anxious great-great-great^n-ancestor so pathologically afraid of floods that he insisted on dwelling on a hill or so fearful of sabre-tooth tigers that he always carried a sharp object that might just be credited for his survival. In fact, Marks and Nesse (1994) argue that eliminating fear would be by no means exclusively positive – and perhaps even the existence of a pathological state of low fear they term ‘hypophobic disorder’. Certainly the lack of fear response, such as that induced by ablation of metabotropic glutamate receptor subtype 7 or interneuronal ablation by inhibition of the Dlx1 gene in the experimental setting or witnessed in the context of pervasive neurodevelopmental disorders both in models and in vivo, has significant evolutionary drawbacks – it is not hard to see how behaviour without fear in a world of danger can quickly lead to reduced life expectancies.
And so we get to the Sensabot, our ‘fearless’ robot. What benefits does his fearlessness yield us? None, I submit, for the reasons below.
Fear is a diverse phenomenon. Cutting it out is neurosurgery with a hatchet.
Fear is a single word (albeit one replete with synonyms) for a number of states of mind that connect somewhere in the human reactions they evoke via the amygdala and thence the limbic system, quite prominently the hypothalamic fight/flight response. It bundles together your fear of spiders (a genuine phobia), your fear of nuclear war (a longer-term anxiety), your fear of wasting your life (an even less acute, more existential perception) and so on. Cutting it all out is doing neurosurgery with a hatchet. A not very sharp one, either.
To put it in the context of the robot: of course, a robot who is not worried about spiders, doesn’t hyperventilate and sweat when handling dangerous substances and doesn’t freeze at the sight of a few zombies emerging from the neighbouring compartment (a risk I doubt Shell needs to envisage at this point, of course!) lacks maladaptive manifestations of fear. These intersect somewhere in the shared fact that they’re either disproportionate to the risk (spiders) or unproductive in the situation (handling dangerous substances and freezing at the sight of a zombie).
But what about other aspects of fear? Fear is a natural way to warn us of existential danger. The Sensabot relies on the human operator to have at least some degree of that, but more autonomous bots will not be able to. Nor do operators act the same when they’re in the cockpit than when they’re piloting an unmanned vehicle. Of course, Sensabots can be replaced far easier than humans, but that’s irrelevant here, both axiomatically and practically (the collateral damage of e.g. an explosion caused by an incorrect action might well be an actual human in the area). Nor does the bot have the neurophysiological advantages that fear – specifically, the cardiac effects of faster movement, better cognitive capabilities and so on. Fear is a reserve, and machines don’t have that reserve.
Fear is best represented as a vector, having both magnitude and direction (example: “I’m very afraid (magnitude) of spiders (direction)”). Different magnitudes help prioritising for immediacy and apprehended risk (likelihood times expected loss). Of course, it is not possible to simply bestow this upon a computer, and there are other methods of prioritising risk, but the great benefit of fear is that it distills signals down into a simple and fast calculation that is remarkably rarely wrong. It does so by considering a current signal in the context of all signals, the signal space of all possible signals, as well as learned patterns and the wider context in which the entire process is taking place. The decision whether to be afraid of something is, actually, quite complex.
This prioritisation is often lampooned when it appears to go wrong, typically when people are afraid of certain rare risks than more frequent ones. Typically, such caricatures of the way human fears work get several things wrong – they reduce the situation to mere probability when in reality, the likelihood of loss, the manner of loss, its impact on others, its impact on your wider community at large and so on are taken into account. That’s why more people fear terrorist attacks than car accidents, even though the latter are much more frequent. Context is everything, and if we learn only one thing from fear, let it be that the evaluation of risks takes place in a very wide context, and with its holistic nature – involving the limbic system, the median prefrontal cortex (mPFC) for memory, somewhat affected by the person’s state of arousal and HPA axis function, mediated by sensory perceptions mixed with our interpretations thereof –, fear is a highly multifactorial response that can be promoted, mediated or inhibited by a number of factors on the way. There is now a degree of awareness in literature that learned and innate fears are differentiated in their propagation pathways (specifically, the involvement of the prelimbic mPFC), indicating again that fear is a single response to different processes reacting to different stimuli. This underlines how fear, rather than simply getting one’s brain pickled in adrenaline, is a complex phenomenon. From an evolutionary perspective, this complexity calls for an explanation – the holistic nature of fear comes, of course, at the expense of the time it takes to trigger release of adrenaline and fight/flight responses. That explanation is, of course, that fear has to accomplish more objectives than merely recoiling, reliably and every time, from a trigger: it has to weigh whether a response is going to put us in more danger, it has to weigh our resources against plans of escape, it has to consider the entire context of a situation, including the past (via the memory activity of the prelimbic mPFC).
Humans are afraid. Their robot co-workers need to understand this.
My learned friend Adam Elkus has made a great point:
@DoodlingData +1. This is common error when thinking about robots. Especially if they have to reason about human theory of mind in co-work
He’s entirely right – if robots and humans work together, robots need to have what is sometimes referred to as the ‘theory of mind’ – an internal concept of an external mind – in order to anticipate and understand their workmates. And humans, well, humans experience fear. There’s no way to getting around that. An interesting consideration here would be a fear related adaptation of Baron-Cohen’s Sally-Anne test, which I will call the Sally-Zombie test. Anne is, in this case, replaced by a rotting carcass reanimated by dark forces. A robot ought to be able to reason about Sally’s reaction to this. Merely predicting a response is unlikely – even in the absence of in-depth knowledge of Sally’s decisions in the past, it will be hard to predict whether she will freeze, run or reach for the nearest object to decapitate her once-friend-turned-zombie with. The robot, thus, ought to understand multiple things here:
- The human emotional context: humans and their fear of their body being taken over, a fear rooted in the human appreciation for autonomy and capacity.
- The human cultural context: zombie movies are a staple of Western culture and at least to Western observers, zombies are unequivocally scary (more advanced models would need to consider cultural differences, e.g. the response Sally would have, had she grown up with a culture, such as Haitian Voodoo or Palo Malombe, where zombies occupy a more complex albeit still fear-inducing position).
- The human personal context: what are Sally’s experiences with zombies? With similar stressors? With the concept of losing a friend to an abomination?
- The human physiological context: given Sally’s state, what is she most likely to experience when her body goes through the motions of fear/panic and the corresponding neural level reactions?
The fact is that the qualia of fear is a uniquely human perception. No machine, however intricate, will ever have the qualia of fear. Whether it needs the qualia, however, in order to do its job is debatable. At the same time, its sheer richness and multifactorial nature, as well as extent of its manifestations, make fear unsuitable to a mere scripted response level of understanding. While some basic features can be easily responded to without having an understanding of fear (“If Joe is around tons of highly explosive material, Joe will have a heart rate exceeding his basal heart rate by approximately 10-25% and experience palmar hydrosis”), most cannot. And that means that robots, to be effective, have to develop something in between the unattainable qualia and the insufficient scripted understanding.
It’s important to understand that this is not merely for robots to understand how their human companions will think, but also to allow the latter to understand how their robot coworker would perceive a situation. A coworker who lacks, say, an ordinary human level of fear might not only endanger his companions, his actions are also likely to be unintelligible to them: why is Steve running towards the madly out-of-control spinning saw blades without any protective equipment?! Fear is so fundamental to being human that it is part of the unwritten set of shared presumptions that help us understand and anticipate each other. This is a system into which robots will, someday soon, integrate themselves.
Do we want robots to be afraid? Some applications, such as the recent proliferation of humanoid, emotionally expressive robots for therapeutic, educational, or general usage purposes certainly rely on at least an understanding of where the display of the physiological-communicative responses to stimuli is appropriate. But that’s not where the story ought to end. In fact, robots that need to have more than trivial ability to reason on their own, especially if they need to do so in a human context. There is much that robots can do better than humans – they don’t feel pain, remorse, regret, doubt, boredom or fatigue. To inject into what one might perceive as almost perfect creatures the maladaptive aspects, too, of human responses to perceived dangers sounds counterintuitive. At the same time, on the large (evolutionary) scale as well as the individual long-term scale, fear is a gift. It is a gift we as humans must consider to pass on to our creations.
|↑1||That’s ‘making your heart beat stronger’ and ‘making your heart beat faster’ in human language, respectively.|
|↑2||Miller, Gregory E., Edith Chen, and Karen J. Parker. “Psychological stress in childhood and susceptibility to the chronic diseases of aging: moving toward a model of behavioral and biological mechanisms.” Psychological Bulletin 137.6 (2011): 959.|
|↑3||Sasaki, Aya, Wilfred C. de Vega, and Patrick O. McGowan. “Biological embedding in mental health: An epigenomic perspective 1.” Biochemistry and Cell Biology 91.1 (2013): 14-21.|
|↑4||Nesse, Randolph M. “Fear and fitness: An evolutionary analysis of anxiety disorders.” Ethology and sociobiology 15.5 (1994): 247-261.|
|↑5||Masugi, Miwako, et al. “Metabotropic glutamate receptor subtype 7 ablation causes deficit in fear response and conditioned taste aversion.” The Journal of Neuroscience 19.3 (1999): 955-963.|
|↑6||Mao, Rong, et al. “Reduced conditioned fear response in mice that lack Dlx1 and show subtype-specific loss of interneurons.” Journal of Neurodevelopmental Disorders 1.3 (2009): 224.|
|↑7||Markram, Kamila, et al. “Abnormal fear conditioning and amygdala processing in an animal model of autism.” Neuropsychopharmacology 33.4 (2008): 901-912.|
|↑8||Consider DSM-IV 299.0, at Associated Features and Disorders, para.1|
|↑9||Said at risk of massive oversimplification. It is WAY more complex than that, of course, and the mPFC as well as other parts of the brain play a significant role. There is an increasing understanding that some of our most fundamental emotions like fear are as close as one gets to global in the brain!|
|↑10||As this song attests.|
|↑11||Corcoran, Kevin A., and Gregory J. Quirk. “Activity in prelimbic cortex is necessary for the expression of learned, but not innate, fears.” The Journal of neuroscience 27.4 (2007): 840-844.|
|↑13||Baron-Cohen, Simon, Alan M. Leslie, and Uta Frith. “Does the autistic child have a “theory of mind”?.” Cognition 21.1 (1985): 37-46.|
|↑14||Buck, Ross. “What is this thing called subjective experience? Reflections on the neuropsychology of qualia.” Neuropsychology 7.4 (1993): 490.|
|↑15||Movellan, Javier, et al. “Sociable robot improves toddler vocabulary skills.” Proceedings of the 4th ACM/IEEE international conference on Human robot interaction. ACM, 2009.|
|↑16||Consider in this field Hashimoto, Takuya, et al. “Development of the face robot SAYA for rich facial expressions.” 2006 SICE-ICASE International Joint Conference. IEEE, 2006. and Itoh, Kazuko, et al. “Various emotional expressions with emotion expression humanoid robot WE-4RII.” Robotics and Automation, 2004. TExCRA’04. First IEEE Technical Exhibition Based Conference on. IEEE, 2004.|