When the average person thinks about AI and robots what often comes to mind are post-apocalyptic visions of scary, super-intelligent machines taking over the world, or even the universe. The Terminator movie series is a good reflection of this fear of AI, with the core technology behind the intelligent machines powered by Skynet, referred to as an “artificial neural network-based conscious group mind and artificial general superintelligence system”. However, the AI of today looks nothing like the worrisome science fiction representation. Rather, AI is performing many tedious and manual tasks and providing value from recognition and conversation systems to predictive analytics pattern matching and autonomous systems.In that context, the fact that governments and military organizations are investing heavily in AI shouldn’t be as much concerning as it is intriguing. The ways that machine learning and AI are being implemented are both mundane from the perspective of enabling humans to do their existing tasks better, and very interesting seeing how machines are being made more intelligent to give humans better understanding and control of the environment around them.
John Fossaceca, APM for AI & ML for Maneuver & Mobility at US Army Research Laboratory (ARL), who spoke at a recent AI in Government event shares some insights as to how AI is being applied on a day-to-day basis as well as where things are heading with autonomous bots and other machines in the US Army.
How is the Army currently leveraging AI?
John Fossaceca: The Army is leveraging AI in many ways, for example in predictive maintenance. AI techniques can help predict when vehicle parts need to be replaced or serviced before the vehicle breaks down. If this can be done well it will save money and increase operational safety. This is being implemented with the Bradley Fighting Vehicle as well as others.
Recommended For YouThe Army has a vast amount of data and many AI and Machine Learning (AI/ML) techniques require large amounts of data. Some programs that leverage data include Project Maven that consumes data from drones and helps to automate some of the work that analysts do. Project Maven leverages some standard AI tools such as Google’s TensorFlow as well as customized tools built internally.
The Army has active ongoing research using AI to enhance autonomous vehicles, electronic warfare and signal intelligence, sensor fusion and augmented reality. AI will improve situational awareness in the battlefield and improve decision-making with programs such as the Joint All-Domain Command & Control (JAD-C2) initiative.
Another area where AI plays a role for the Army is in talent management. The Army’s AI Task Force (AITF) has an initiative to use AI
Army Futures Command AI Task Force
Army Futures Command AI Task Force
to identify the competencies and attributes that lead to successful performance that can then be used to find potential candidates for positions in the Army.
At the Combat Capabilities Development Command’s Army Research Laboratory (ARL), Artificial Intelligence is considered to be a primary research area. ARL is the Army’s corporate research laboratory and has many initiatives that leverage Artificial Intelligence. For example the essential research program entitled, Artificial Intelligence for Maneuver and Mobility (AIMM) is leading the way for how the Army will imbue the Next Generation Combat Vehicles (NGCV) with the ability to operate off road without the need for being supervised by a soldier with a remote control radio. These next generation intelligent vehicles will be able to reason about specific situations, environmental conditions and make decisions about the best action to take while keeping soldier teammates informed and improving overall situational awareness. There are many other essential research programs (ERPs) at ARL that also leverage AI methods and all of these ERPs are producing innovations that will greatly benefit army operations in the future.
In the near term, the Army is using AI to leverage inputs from multiple sensors in order to build an accurate picture of battlefield threats and speed up the targeting and decision making process in Project Convergence, an initiative led by the Army Futures Command.
What are some challenges in the Army when it comes to AI/ML adoption?
John Fossaceca: Commercial AI relies on vast computing resources and large amount of data including cloud computing reach back when necessary. Battlefield AI, on the other hand, must operate within the constraints of edge devices: Computer processors must be relatively light and small with potentially constrained communication bandwidth under adversarial conditions.
In Army applications often there is either not enough training data or the data is corrupted or noisy. Operational environments tend to be dynamically changing and sometimes unstructured with damaged roads, buildings and infrastructure. There is heterogeneous data from many sources, sometimes this data is deceptive or influenced by adversaries.
Today’s AI techniques tend to be brittle and can break down even under ideal operating conditions. These methods are very limited in their ability to reason, especially in real-time. There are some deployed systems that tout AI capabilities that are limited to hard-coded rules and lack the ability to reason and infer from inputs from sensors and other systems and do not provide enhanced situational assessment.
Many of the AI approaches depend on “supervised learning” (e.g. deep learning) and these techniques create massive models, often with 10 to 100 million parameters learned in a “batch-based” mode on powerful computing infrastructures. The Army needs alternatives to these offline and time consuming training methods.
Ultimately, current systems are not able to operate autonomously, and require constant human attention, intervention and manual control. Back in 2018 we were looking at learning from feedback where a human observer would simply provide a positive or negative signal to the intelligent agent and we demonstrated that we could drastically reduce the learning time by orders of magnitude. We are extending this research to Learning from Demonstration which I’ll discuss soon.
As our research progressed we realized that we need a way to interact and communicate with intelligent agents in a natural way. Beyond just natural dialog and grounding, a lot of issues crop up due to a lack of shared understanding of the world and “common sense reasoning”. These shortcomings are being addressed through several research programs in AIMM’s second line of effort – Context Aware Decision Making.
How is the Army working towards getting their data in a usable state for AI/ML?
John Fossaceca: There are many data collection and labelling initiatives being worked on by the Army and across the DoD preparing data for use by AI algorithms. For example, Project Maven has a lot of videos from military drones. Sometimes labelling is done through crowdsourcing techniques depending on the level of classification. Other initiatives include ARL’s work to internally collect data from various locations and with research partners to curate and label data from a variety of terrains. ARL has a Robotics Research Collaboration Campus (R2C2) in Maryland where data is collected and autonomous experiments are conducted.
In addition to project Maven, there are several efforts across the DoD for intelligence analysis using state of the art tools. Many of these projects focus on detecting specific objects in images using deep learning methods and each of these programs requires large amounts of data be cleaned, curated and labelled in order to be useful. These efforts also require an AI pipeline that consists of storage, algorithmic toolkits, computing resources, testing and deployment tools. Often data format standards are developed to ensure consistency between experiments and tests and provide users with a familiar environment. These data repositories need to be cataloged and be accessible to users as well as have useful descriptions of the data contained within them. There are some efforts to standardize this access information across several databases to make it easier for the intelligence community to use.
How is the Army leveraging AI-enabled autonomous vehicles for Maneuver and Mobility?
John Fossaceca: In the Army’s Robotic and Autonomous Systems (RAS) strategy, General Daniel B. Allyn, Vice Chief of Staff states, “The integration of RAS will help future Army forces, operating as part of Joint teams, to defeat enemy organizations, control terrain, secure populations, and consolidate gains. RAS capabilities will also allow future Army forces to conduct operations consistent with the concept of multi-domain battle, projecting power outward from land into maritime, space, and cyberspace domains to preserve Joint Force freedom of movement and action.”
According to the RAS strategy “Effective integration of RAS improves U.S. forces’ ability to maintain overmatch and renders an enemy unable to respond effectively. The Army must pursue RAS capabilities with urgency because adversaries are developing and employing a broad range of advanced RAS technologies as well as employing new tactics to disrupt U.S. military strengths and exploit perceived weaknesses”
In order to accomplish the vision laid out in the RAS strategy, autonomous vehicles will need to ensure “freedom of maneuver” while decreasing risks to soldiers. This will require collaboration between humans and machines that is autonomous. Vehicles will be teammates for soldiers in the battlefield rather than just another piece of equipment. These “integrated human-machine teams will allow forces to learn, adapt, fight and win under uncertain situations.”
AI is one of the key enablers for these intelligent autonomous systems. These systems will be able to deal with near peer or peer adversaries who can operate at fast speeds by allowing our forces to make decisions more quickly. The Army will also have to contend with the fact that our adversaries will also be using their own autonomous systems. With more autonomy, robotic autonomous systems will be less dependent on communication links that are often unreliable in battlefield conditions due to jamming or capacity issues.
In terms of priorities, the RAS strategy calls for near term improvements in situational awareness and helping to reduce the physical load on soldiers. In the midterm, “automated convoy operations” will not only help with sustainment but will protect soldiers. In the longer term, autonomous vehicles will execute advanced tactical maneuvers and “will increase capabilities within brigade combat teams.”
What are some unique environmental challenges that impact the research that goes into autonomous vehicles and equipment?
John Fossaceca: In addition to complex terrain and unstructured environments where the Army operates, the environment often consists of adversaries and these adversaries may be unpredictable. The Army has research that focuses specifically on so-called “tactical behaviors”, that is, what are the specific formations that the autonomous vehicles should utilize? How can an autonomous vehicle achieve a position of advantage over an adversary? How can an autonomous vehicle operate without being detected by an enemy force? The Army has done research in autonomous subterranean exploration as well as and in order to operationalize autonomy, next generation combat vehicles will need to be able to reason about the possibility of all potential routes, even water crossings.
How does the ARL’s research in autonomous vehicles differ from what industry is doing?
John Fossaceca: Often, in Army contexts because large amounts of militarily relevant, labeled data are not available so a very important research area ARL is pursuing are AI algorithms than can learn with far fewer examples than traditional supervised approaches. In concert with these the Army has developed some unsupervised approaches for things like scene segmentation which can use self-labeling methods. However, such methods still require a lot of computing power and it is challenging to do in real-time on the autonomous vehicle. To help address this problem, the Army has several computer scientists who specialize in computer architecture and algorithms to take advanced state of the art methods and make them work within the processor size and power constraints of Army autonomous vehicles.
The Army has unique technical challenges that the commercial sector is not addressing. Autonomous vehicles generally do not operate in an environment that is contested in all domains. Certainly there are people, obstacles and sometimes unexpected events, however, military operations occur in very uncertain environments, complex and dangerous terrains which may be filled with adversaries and other dangers.
The first instantiation of this will be tele-operated and as the Army operates these vehicles, we will learn how to employ robots in the battlefield. This will inform the autonomous behaviors that we need to develop. Ultimately, Next Generation Combat vehicles will have the capability to learn in the field, adapt to the current situation, reason and act effectively in support of the Multi-domain Operations mission.
What is a unique insight into your AI challenges that others might be interested to learn?
John Fossaceca: Recent Army research has found success with deep reinforcement learning techniques that leverage human demonstration and feedback. Newer methods have been successful in greatly reducing the time it takes to train a system on new tasks. Other research that involves learning from human demonstration is showing early promise and utility for battlefield retraining with the potential for real-time learning using limited examples. These techniques appear to allow for transfer learning, that is, learning under one set of conditions and operating under a new set of conditions without the need to for training from scratch.
How does the Army envision the warfighter and battlefield of the future?
John Fossaceca: The Army’s vision of the future battlefield will have an unmanned formation several kilometers in advance of a manned formation. One goal is to have the autonomous systems do area and route reconnaissance to find or make contact with the enemy while providing stand off for soldiers.
How important is AI to the Army’s vision of the future?
John Fossaceca: AI will be a critical enabler for future success in Multi-domain Operations. According to the former Secretary of the Army and current Secretary of Defense, Mr. Mark Esper, “…if we can master AI … then I think it will just really position us better to make sure we protect the American people. Winning on the future battlefield requires us to act faster than our enemies while placing our troops and resources at a lower risk….Whoever gets there first will maintain a decisive edge on the battlefield for years to come.”
The current Secretary of the Army, Mr. Ryan McCarthy has stated that cloud based technologies and capabilities are key in order to “maximize AI”. Mr. McCarthy wants to see the cloud infrastructure put into place as a driver for AI progress. According to Mr. McCarthy, this will be critical for decision-making in the battlefield.
What is the Army’s perspective on ethics and responsible use of AI?
John Fossaceca: The Army and DoD as a whole is concerned with AI Ethics and last October a draft of “Recommendations on the Ethical Use of Artificial Intelligence.” These rules will apply across the U.S. military. The U.S. military will have humans that will be in control of all AI enabled systems.
The Army’s AI Task Force has an ethics officer who helps to inform AI ethics policies. Per Secretary of the Army, Mr. Ryan McCarthy, “A system can crunch the data very quickly and give you an answer, but it doesn’t have context,” he said. “Only a human being can bring the context to a decision.”
What are you doing to get an AI ready workforce and war fighter? Are you providing training and education around AI?
John Fossaceca: ARL and the Army offers many opportunities for students to do internships as well as SMART Scholarships that help students pay for education and in exchange the student will work for the Army for a period of time. ARL also hires new doctoral graduates as PostDocs and brings them in to do cutting edge research. Eventually some of the PostDocs will become employees. Because Artificial Intelligence is key competency area, the Army is increasingly hiring scientists and engineers with this expertise.
What are you doing now to train soldiers to make them more comfortable working alongside autonomous systems and robots?
John Fossaceca: Since the types of autonomous systems we are talking about are still under development, we use simulation in our training environments to help soldiers get comfortable operating with autonomous systems. The Army is early in process but there are some ongoing initiatives such as the Reconfigurable Virtual Collective Trainer (RVCT) including both ground and air platforms that provides the ability to rehearse missions with simulated data.
Many of the training efforts focus on realistic simulations of intelligent semi-autonomous and autonomous systems providing soldiers an immersive training experience. Soldiers train against virtual opponents in this Synthetic Training Environment (STE). These virtual opponents are imbued with intelligent behaviors that have a certain unpredictability to simulate adversaries as well as a reasonable level of cognition based on also leveraging state-of-the-art artificial intelligence with realistic environments.
At the foundational research level, ARL has used soldiers to interact with autonomous prototypes to learn how soldiers speak and what commands they tend to use. This has also help some soldiers learn how autonomous systems behave. In fact, as soldiers train with autonomous systems, they tend to adapt their language over time to more effectively communicate and control such systems.
What AI technologies are you most looking forward to in the coming years?
John Fossaceca: We are making advances in using artificial intelligence for reasoning about the environment and being able to recommend specific courses of action to soldier teammates. This will represent moving beyond narrow AI, where autonomous agents can do very specific tasks well, to being able to adapt to new, never before seen situations. They will determine what actions are possible and the probability of success for each of these actions. This is not “general AI”, that is, AI that can reason at a level close to human beings. What we envision in the future, is the ability for autonomous systems to do sophisticated reasoning about a given situation, make complex decisions and anticipate what the outcomes might be to ensure mission success.