The future of warfare will be fast and unrelenting. An automated kill chain from sensor-to-shooter is already in our sights. Hypersonic weapons, loitering drones, and munitions that alter course in flight are changing the way we fight. While humans seek to retain overall decision‑making authority, delegating only what is too complex or quick for us to manage, future Army operations should push objectives and tasks to artificial intelligence and autonomous agents for execution. To do so, however, an autonomous agent must be able to sense its environment to learn relevant rules and thereby ensure that it is executing the intent of its mission.

The plot of Karel Čapek’s 1920 play R.U.R., where he coined the term “robot,” addressed the question of whether these robots could self-actualize—to become like humans, to yearn, to dream. Such a conception of robots is in the minority, in contrast to the dominant views today, which generally take a functional perspective: robots are either switched off or industrious—a binary existence. Absent an accident or maintenance, robots have no downtime. In R.U.R., one character, in describing the invention of the robot, says “Man is something which . . . feels joy, plays the violin, takes a walk, has the urge to do all sorts of things which are actually quite useless.” With AI or autonomous agents, the productivity that comes with not paying attention to these useless distractions is paramount.

Ignoring the Commonplace

A robot traverses a long corridor of closed doors in a hotel. It reaches the end of the corridor. The end of the corridor, after the monotony of closed doors, is a novelty. It is different and this is something the robot can be programmed to detect. For the sake of efficiency, much of the detail of the corridor can be discarded—what matters is how long it took to reach this landmark. Other sights may take the robot’s attention—an elevator door opening, a tray of discarded room service dishes, a fire extinguisher. These details may hold relevancy in the context of what the robot has observed. In human terms, they’re interesting.

While one may scoff at a tray of dishes being interesting, remember the hallway of one floor of a hotel encompasses this robot’s entire existence. As the robot gains experience, trays, elevators, and fire extinguishers may also become commonplace—that is, as the robot experiences more of the hotel, they will lose their novelty in relation to the swimming pool, the breakfast room, the rooftop bar—certainly anywhere noisy, unpredictable people congregate. Because these lesser landmarks lose significance, they can be pruned from the robot’s model of its world, just as humans view things like room service trays and fire extinguishers to be so commonplace that they become irrelevant and are forgotten. If someone asks about a hotel stay, the description given in response may include the pool, but it is unlikely to include when the elevator doors opened.

Failing, Falling, Learning

Imagine a well-meaning guest has propped open the fire door to the stairwell, and our intrepid explorer wanders through and falls down the steps. Interesting! This is certainly a novelty. Yet, how useful is the fall alone, the actual novel event? AI succeeds at identifying patterns, but no pattern exists for single occurrences.

After repair, the robot again crosses into the stairwell, tips over the edge, and falls—it calculates the similarity to the previous fall too late to prevent it happening again. The novelty of the second fall is also lower than that of the first. Learning does not result from recalling the most novel events. The robot must link less-notable events (the open stairwell door, crossing the stairwell threshold, stairs appearing) with the novelty of falling down the stairs. Linking two events results in associative memory.

Machine learning is an AI tool that uses statistical calculation to make probabilistic predictions. A key feature in this approach is that there is no need to explicitly program the relationships—machine learning infers them. Take, for example, recommendations on a streaming service. If other people who watched the same movie you enjoyed also liked a movie you haven’t watched, then the service may infer your enjoyment of the other movie. It doesn’t know that for sure—it’s making a prediction based on what it knows about user viewing habits.

The goal of machine learning is to create a vector space where the distance between elements represents a relationship. One vector may be that both movies have the same director. Other vectors may be that the movies feature the same actor, setting, or general vibe. Perhaps the same person liked both, and any additional people who liked both movies strengthened that connection, closing the distance between them in the vector space. The stronger the relationship, the more likely that if you liked the first movie, you’d like the second. A deep-learning approach uses networked layers of these correlations to optimize prediction, but requires large amounts of data to do so effectively.

Common examples of patterns depend on visual properties—stripes, gingham, polka dots, sequences. However, relationships can also be established through proximity in time. Russian scientist Ivan Pavlov is known for his experiments in classical conditioning. Although his primary interest was in salivation rates and digestion of his canine subjects, Pavlov noticed that the assistant charged with feeding the dogs elicited a salivation response even when he appeared without bearing food. He further abstracted this cue by playing a tone before the dogs were fed until the dogs associated the tone with food and would begin to salivate when they heard the sound alone.

This relationship works in a different way than the example with movies—a dog who likes food doesn’t also like tones, of course. Time is the factor here. If Pavlov played a tone and then fed the dogs hours later, the two events would not be associated, no matter how consistent. Our intrepid hall-crawling robot needs these types of correlation with the sequencing of events to associate crossing the threshold into the stairwell with falling down the steps.

It’s important to note that while correlation does not mean causation—Pavlov’s tone did not cause the food to be delivered just as crossing the stairwell threshold does not cause the robot to fall down the stairs—causality isn’t necessary when establishing a statistically relevant relationship. The value of the relationship is in its ability to facilitate a prediction.

Pavlov’s conditioning took place over multiple iterations—playing a tone one time before food is delivered is insufficient to establish the association. How many times will our robot fall down the steps before it can create a rule that preempts future falls?

Dream a Little Dream

In the title of his 1968 novel, later adapted into the film Blade Runner (1982), Philip K. Dick asks: “Do androids dream of electric sheep?” On its face, the idea is absurd—a sleeping robot. Authors almost universally portray robots as hyper-vigilant and unrelenting—just try getting C3PO never to tell you the odds. Surely one of the advantages of an android is that sleep is useless to it.

Sleep is a liability for creatures as soft and tasty as humans. If humans have evolved into such a liability, there must be a benefit to balance the risk. There is evidence that sleep in general and dreaming specifically provides a state for association. In our dreams, we can revisit memories and make connections in ways not possible while consumed with the activity of consciousness. This would allow strengthening associations through repetition without having to repeat the physical event.

While computers are ideal for multitasking, they have performance differences in trying to develop associations as events unfold versus processing and pruning after the fact, once removed from the situation (like humans, robots benefit from hindsight even without rear-facing sensors). Despite the apparent benefits, when dividing human and machine tasks for optimization, one would expect resistance to signing robots up for nap time.

Perhaps we should look to the skies and the seas for a solution. The unihemispheric slow-wave sleep model of birds and water mammals, where the hemispheres of their brain take turns sleeping, could provide a lesson. If our hallway robot does a mind meld with another computer—offloading a copy of its data—this second agent can conduct the postprocessing independently, identifying and strengthening associations using an approach known as one shot or few shot learning that builds on context and prior experience to make fresh insights. In the morning, our sleeper can deliver fresh algorithms with the same associations our primary bot would have gotten had it dreamt itself. Machines with each of these roles—sensing and sensemaking—can be optimized using machine learning for their particular tasks.

Further, dreams can efficiently build associations without the risk of physical danger that comes from making mistakes in the real world. This is key in a tactical situation, where scenarios are complex, data is hard-earned, and we only get one opportunity to get things right.

The speed of the battlefield will continue to increase, but the effectiveness of our approach is not only measured in microseconds, but also in the ability for our frontline resources to adapt to novel circumstances and unique events. Big-data machine-learning solutions assume not only the availability and quality of data, but also that it is predictable within certain constraints (for example, consistent inputs). Meanwhile, the battlefield is a premier example of an uncontrolled environment. We cannot rely on only a single type of learning for all applications.

In a quest to optimize efficiency, the creator of robots in Čapek’s R.U.R. “invented a worker with minimal needs. He just simplified it, discarded all the various optional devices unrelated to work.” People often quickly discard what seems not to have direct value to an endeavor—the context that makes an object or an event significant in comparison, the dreams that seem too fanciful to relate to reality. Artificial agents are supposed to be a better version of ourselves, but to be successful, we still need to establish which parts of us are good.

Thom Hawkins is a project officer for artificial intelligence and data strategy with US Army Project Manager Mission Command. Mr. Hawkins specializes in AI-enabling infrastructure and adoption of AI-driven decision aids.

Troy Kelley is a retired researcher from the US Army Research Laboratory. Mr. Kelley spent thirty years researching robotics, cognition, and AI. Mr. Kelley was the principal developer of the Symbolic and Sub-symbolic Robotics Intelligence Control System, which was a robotics research and development platform for the Army, and the Modeling and Integration of Neurological Dynamics with Symbolic Structures program, which was a program to model cognition on the Army’s super computers.

The views expressed are those of the authors and do not reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense.

Image credit: Airman 1st Class Luis A. Ruiz-Vazquez, US Air Force