Close Encounters of the Robotic Kind: A Glimpse of Autonomy
Earlier this summer, people around the world were gripped by the story of the young boys from a Thai soccer team who had been trapped inside of a cave after a flood. We watched with fascination – and a healthy dose of apprehension – as plans were formulated to rescue the boys, whose situation was becoming increasingly perilous as floodwaters rose and oxygen inside the cave began to deplete. Thanks to the incredible efforts of some of the world’s best cave divers, all of the boys and their coach made it out of the cave alive in what many would describe as nothing short of a miracle.
Use of the term “miracle” may be hyperbolic, but it is nonetheless fitting – the exploration of submerged caves is a remarkable feat of mental and physical discipline. Even this successful rescue mission was not without its casualties, as an ex-Thai Navy Seal lost his life laying down oxygen tanks in preparation for the rescue. With the story of the Thai cave rescue hot on people’s minds, I came across a common theme in my social media feeds: everyone is thankful for those cave divers, but what on earth motivated these people to start venturing into dangerous sea caves in the first place?! It is a fair question, but I’d argue the answer is underwhelmingly simple: humans are a curious animal. Cave diving may pose a thrilling adventure, but as a diver myself and the daughter of a cave diver, I wouldn’t necessarily describe the dive community I know as a population of adrenaline junkies like I’ve seen so many speculate. The preparation just to enter the water is time-consuming, painstakingly methodical, and I dare say even boring to someone who craves fast-paced action. If anything, the unifying theme among the divers I have met is a curiosity reminiscent of many of the scientists I have met during my own career in laboratories. It is this thirst for knowledge and unique experience that I posit drives otherwise rational human beings into unforgiving environments devoid of light, air, and even wifi.
History is filled with examples of brave people venturing into the unknown, and we have benefitted greatly as a collective from the knowledge that those individuals gather. Whether it is the unknown of a cave, or the ocean, or outer space, we love the idea of sending someone out into the field so that they may regale us of their journey and relay valuable data. There are some places, however, that it seems rather impossible for us to send an explorer with currently available technology. We may have set foot on the moon, but even the most daring among us isn’t going to make it to one of Jupiter’s moons and come back to tell the tale. Yet there are things in these unreachable places that we desperately want to investigate. Jupiter’s moon Europa, for example, may be capable of sustaining life beneath the ice of its vast ocean. For these destinations, a promising solution for our exploratory hurdles may lie in the development of autonomous exploration vehicles.
There are some immediately evident benefits to sending a robot into the unknown rather than a human. It can’t die, for starters. But human explorers would be out of a job by now if a remotely controlled robot, or even one with a pre-programmed map or set of commands, could do everything that a human can without the risk of death. The fact is that most robots lack one of the most important aspects of humanity for the job of exploration – autonomy. Most robots are not capable of mitigating a problem with quick decision-making or navigating an unknown landscape, both of which are human abilities critical to the pursuit of exploration. Historically robots that do not require active human control have been programmed to pursue a specific task and can do so pretty well under the ideal conditions in which it was designed to function. Exploration of the unknown, however, doesn’t tend to come with a map or detailed set of expectations, and we can’t send the average robot all the way to Jupiter if it handles an unexpected challenge the way a Roomba handles an extension cord.
The challenge of unmanned exploration is one driving force behind the development of autonomous robots – autonomous here meaning a robot that can perform tasks on its own. Technically by this loose definition, a Roomba is autonomous because it can perform the task of cleaning a room on its own. Within the extremely limited confines of a specific task, this robot is autonomous. There is a spectrum of autonomy, however. The Mars rovers, for example, would be considered much more autonomous than a Roomba in that they are much more sophisticated and capable of accomplishing a variety of complex tasks. Even those rovers need human help once in a while, though, when they get stuck in a sand dune or otherwise tied up by the planetary equivalent of an extension cord. Beyond the realm of robotics lie definitions of autonomy more akin to that which a philosopher or even a neuroscientist like myself might use: the capacity to make an informed, uncoerced decision. A politician might define autonomy as the capacity to self-govern. The Merriam-Webster medical dictionary defines it as the state of being independent, free, and self-directing. These definitions are all similar, but the nuances in the definitions are reflective of the nuances with which we tend to view the concept of autonomy in general – it doesn’t always mean the same thing.
I can appreciate the wondrously complex technology that must go into building autonomy from scratch. But metal gears and lines of code have never been major characters in my life story, as it were. I am more intrigued by these robotic creatures because I want to know what autonomy looks like when people have the chance to implement it in their own vision. How does an autonomous robot make decisions? We don’t even fully comprehend how we make decisions. Can a machine learn more efficiently than a human, unburdened by self-doubt or anxieties? Does it want for creativity where an inelegant code may be too rigid? At what point do we enter full sci-fi territory and glimpse the birth of mechanistic self-awareness? Do autonomous behaviors start to form a personality? Would that lead to an autonomous vehicle too stubborn to consult external sources for directions? And is it going to happen to the self-driving ride share car that I will doubtless find myself in someday?
Alas, I no longer have to lie awake at night pondering these questions, for I have an insider’s account of what it’s like to work with an autonomous robot.
My father Joe, the aforementioned cave diver, was asked to assist with the testing of an autonomous underwater vehicle (AUV) named Sunfish. This AUV is being developed by Stone Aerospace, and it will have to meet much higher standards of autonomy than a Roomba to achieve its goal of going to Europa one day to search for signs of life. Joe served as the head of communications among a crew of experienced cave divers who were asked to accompany Sunfish on a journey through an underwater cave system in North Florida’s Peacock Springs State Park. The goal of the week of testing was straightforward: Sunfish would have to travel from the entry point at Peacock I to a specific location known as Pothole Sink, and then back again, entirely on its own. Testing of its specific features, like sonar and ability to generate 3D maps, had all been done before; this time the focus was on autonomy.
Testing would begin with two or three divers going into the water with Sunfish, which at this point is fiber optically tethered to a team of engineers on land so they can monitor the running code in real time. They could receive data through this tether, but could not control Sunfish. Joe would wear a full-face diving mask to communicate Sunfish’s behaviors to the engineers. Sunfish would enter the environment, spin to create a full 3D map, assess navigational options, and proceed. There wasn’t a programmed detailed map, since there won’t be one on Europa. It would have to know that it had reached its destination in another way – by sunlight, perhaps, or other life.
There was some modification to be made the first few days, to be sure. They spent several hours on the first day working to achieve neutral buoyancy with Sunfish. A diver seeks to be neutrally buoyant – neither floating to the surface nor sinking to the bottom – so as to swim smoothly without bouncing off the ocean floor or into a cave ceiling. Likewise, Sunfish would have to find neutral buoyancy so it wouldn’t need thrusters running constantly to keep it off the floor, and it did find this balance. It also needed some help interpreting these foreign surroundings. Little pockets of air can create glistening mirror-like effects in the water, and pillar-esque structures in the cave proved similarly confusing. “Sometimes it just stopped and stared,” Joe said. But every time that happened, the code was fixed overnight and the problem didn’t happen again the next day. In the grand scheme of things, the Sunfish needed way less human interference than you might assume. Joe recalled that the programmers seemed pleased and never stumped by any challenges that arose.
Those initial dives with Sunfish went swimmingly. Little problems arose and were solved as quickly as they had emerged. Sunfish needs to be as good and autonomous as the real human deal if it’s going to space, though, so I pressed my father on the issue. What, as far as he could tell by spending time with it, were Sunfish’s biggest limitations? Is there a uniquely human strategy that Sunfish just can’t pull off? “I don’t know if it has a limit,” he said. We discussed what a diver thinks and feels in the moment, and the decisions that need to be made. When Joe is in a cave, he relies on his senses in conjunction with his own mental map and reacts based upon that combination of information. You have to evaluate multiple scenarios simultaneously. Does this arm of the cave seem too restricted? Is the flow of the water along this route too strong to manage? He paused as he went through this process, and then suggested that this must be similar to what Sunfish is doing; perhaps our thoughts and decisions are just like lines of code, but happening so quickly we don’t perceive it. It’s a comparison all too familiar to the increasingly popular field of computational neuroscience, and it’s at the heart of efforts to develop sophisticated artificial intelligence. Can we write a code for smart decision-making? Or perhaps more human attributes, like empathy or consciousness? Where precisely does the comparison between brain and machine break down? And does it break down because the brain is a superior entity, or because we just haven’t learned enough to replicate it yet?
Joe did highlight one consequential difference between diver and Sunfish, though it wasn’t exactly a point in human’s favor. “Every human has a point at which they’ll panic.” Even divers who are regularly out in the water will have their moments. It’s bound to happen even more frequently in these unknown places. With panic comes mistakes, and in an unforgiving environment like a submerged cave, mistakes can mean death. To that end, Sunfish gains a major advantage. It can’t panic, no matter how dark and claustrophobic that cave gets, and its mistakes are unlikely to be life-threatening.
After a few days of testing, it was time for the main event. Sunfish would be tasked with a journey from Peacock I to Pothole Sink and back again with absolutely no human interference. That return trip is no easy feat, according to Joe. He likened it to going to the mall, running your errands, and then trying to find your car at the end of your trip – just because you traveled that route before does not mean it all will look the same on the way back. The dive was first mapped by man in 1967 when Sheck Exley – pioneer cave diver instrumental in the exploration of many caves in Florida and Mexico, as well as math teacher and dude with an awesome name – completed the journey solo. It was a major dive at the time, though now it would be considered a basic dive for a novice cave diver. Now fifty years later, Sunfish would try to replicate that dive, and it did so flawlessly. “It took exactly the route that I would have,” Joe remarked. It successfully recognized its arrival at Pothole Sink, and subsequently navigated smoothly back to its origin point. It was the first ever fully autonomous cave dive by an unmanned vehicle.
We may not yet have a consensus definition for what it means to be truly autonomous, but insofar as that definition includes observation, decision-making, and execution of the appropriate action entirely devoid of interference by a third party, we have made unbelievable progress in the automation of autonomy. We see that progress all the time in the news, with a range of success and failure, from Mars rovers to self-driving cars. These technologies have wildly expanded our capacity to explore our planet and its neighbors, and also modified our daily lives by freeing us from having to perform some of life’s more mundane tasks so that we may pursue things more exciting that vacuuming the living room. It is not yet clear how far we will be able to extend the development of autonomy in machines, or the extent to which we may be able to replicate human phenomena with the right combination of hardware and software. But Joe looks at the story of Sunfish and Peacock Springs as a larger arc that could inform this question. Fifty years ago that dive was a feat of human skill and endurance, and today that destination is but a waypoint on a longer dive. Today that dive was a feat of autonomous exploration, and fifty years from now that destination would be just a waypoint for the next generation of Sunfish. The next stop on that journey? For Sunfish, “it’s off to Europa!” Joe exclaims. For the mechanization of autonomy, the wheels of progress move only forward – the destination depends only on where we steer them.