Can We Trust Artificially Intelligent Machines to Make Decisions in Space?
Yes. I mean no. Well occasionally. Sometimes? Maybe sometimes.
There’s no end to the claims of what Artificial Intelligence (AI) can do, and that’s especially true in space applications. In its loosest definition, AI refers to any kind of ‘intelligence’ exhibited by a machine, and intelligence can mean a lot of things: logic, reasoning, planning, problem-solving, information processing—the list goes on. In this way, AI encompasses a near limitless set of methods, many of which are extremely general and abstract. So which methods are relevant to the exploration of space?
Last year, the European Space Agency (ESA) published a report covering how they’re using AI. Projects ranged from detecting and tracking marine litter using satellites, to mapping the geometry and gravity of space rocks, to molecular modeling of synthetic lattice materials. These diverse projects are each important and exciting in their own right, but in this article, I’m chiefly concerned with how AI can be used to imbue spacecraft and spaceborne machines with autonomy.
Trustworthy artificial intelligence/autonomy is uniquely important in the context of space missions, as the need for making critical decisions independently is greatly amplified. Without this ability, a significant communications overhead is incurred in missions taking place far from Earth (it can take up to 21 minutes for a message to travel one way from Earth to Mars). This reduces efficiency, but also reduces a machine’s ability to react to sudden unforseen circumstances. Obviously, we can only take advantage of this critical decision-making ability if we can be confident in the system’s ability to act reliably and predictably. Hence the importance of trust.
Complex problems, complex solution
In my opinion, one of the most exciting techniques in this area is Deep Reinforcement Learning (DRL). There are two parts to DRL: the Deep Learning (DL) part and the Reinforcement Learning (RL) part…
DL is a branch of AI, where we’re concerned with artificially modeling the cognition of living systems by using algorithms that are inspired by biological brains. Via this act of bioinspiration, we hope to replicate the intelligent capabilities of living systems, computing complex behaviors in a manner similar to how a human does—the only difference being the physical implementation (the hardware on which we carry out those computations). In the world of AI, the most common bioinspired computation model that we use is certainly the Artificial Neural Network (ANN), which uses sequential layers of simple processing to build up to complex behaviors (though there are many other exciting deep learning architectures floating around)
The RL part is simple; it just means that we’re concerned with how our machine should operate in order to get the best outcome, and that this operation can be sequential in nature. For example, in space exploration, RL could be used to navigate a rover across the surface of Mars. The rover learns to make decisions based on feedback from its environment—such as avoiding obstacles, conserving energy, or maximizing the scientific value of its observations. Over time, the rover refines its actions to achieve better outcomes, such as successfully traversing difficult terrain or efficiently reaching a target destination.
More than just a buzzword
So in RL we learn to make the sequences of decisions that achieve an optimal outcome, and in DL we leverage bioinspired computation models to learn like humans do. Perhaps unsurprisingly, this has been used to great effect—many of the AI breakthroughs that you’ve heard of are DRL systems:
Agent57 by DeepMind: Used DRL to learn how to control an Atari 2600’s controller inputs based on visual sensory inputs.
AlphaGo by DeepMind: Used DRL to defeat a world champion in the complex game of Go by mastering learning strategies from experience.
Self-Driving Cars (Waymo & (maybe) Tesla): Use DRL to navigate real-world driving scenarios, continuously improving autonomous driving capabilities through data-driven learning.
ChatGPT by OpenAI: Uses DRL to learn from human feedback to optimize conversational responses, making interactions more helpful and aligned with user expectations.
And in space? DRL applications there are still nascent, with many investigations being limited entirely to simulation, or only being ‘tuned’ on real data. A recent paper in Annual Reviews in Control discussed how “policies for spacecraft control applications are usually trained in a simulated environment”, and continued to list many different applications of this technology, including guidance & control systems, attitude control, spacecraft maneuvers, orbital control, and on-board decision making for spacecraft (check out the paper for more).
It’s not all sunshine and roses
Let’s take a moment to observe that there’s a trade that we’re making here: using DRL improves performance and autonomy, but it does so by leveraging bioinspired computational models that we can’t fully understand and interpret. This makes sense; we need human-like complexity to achieve human-like autonomy, reasoning, information processing, and so on.
Put simply: complex AI models are capable of making optimal decisions, but we don’t always know exactly why they arrive at those decisions.
Obviously, more accountability is needed in the space domain, where our choices decide 1) the wellbeing of expensive craft, 2) the safety of human crews, and 3) the outcome of important scientific missions.
Thankfully, there are a few approaches that can reduce uncertainty:
In missions where the time delay would not be significant, human operators can supervise and ‘OK’ the decisions of artificially intelligent controllers before they are enacted (i.e., keep a ‘human-in-the-loop’).
Where possible, techniques from the field of Explainable AI (XAI) can be used to provide explanations pertaining to the decision processes of AI models (e.g. locally interpretable model agnostic explanations).
For simple enough tasks, the optimal sequence of decisions can be computed up to a known certainty or bounded level of performance, using well-established algorithms (e.g. convex optimization, which was used in SpaceX’s Falcon 9 first stage landings).
For complex tasks where behavior cannot be guaranteed, it may still be possible to statistically analyze the performance of an AI method ‘in the aggregate’ through thorough testing (this approach has been used at Waymo to attain confidence in collision avoidance and performance sufficiency).
How I think about this problem
So, can we trust artificially intelligent machines to make decisions in space? In my opinion the answer is a non-committal (sometimes) yes. It really is situation dependent. In my personal experience, a dependable answer should consider these points (note: these points are relevant in any safety critical situation domain, not just space):
Is the task simple enough to be solved with established algorithms that can guarantee the machine’s performance, or is it complex enough to require sophisticated approaches like DRL? Keep it simple!
Is the AI component actually in charge of anything critical? There are many cases where the AI element may be compartmentalized and unable to cause a meaningful failure.
Is full autonomy actually required? Must decisions be made so rapidly that it would be impossible for a human to be in the loop?
If we do need to use advanced AI techniques (e.g. DRL), then can we:
Gain insight into the AI’s decision making process using bespoke XAI methods that are relevant to the task and model being used?
Quantify our uncertainty statistically by running a thorough battery of tests?
Now let’s go build some intelligent machines. That we can trust. In space. 😎.
—IRW
Great article! Interstellar reference is greatly appreciated.