Artificial Intelligence for Space Exploration
Artificial Intelligence (AI) offers huge potential to augment current and future space exploration. There are a number of applications of AI that could drastically improve future space exploration missions, from greater planetary navigation and optimised mission operations, to analysing astronaut biometric data and enhancing knowledge discovery.
This week two world’s collide: it’s World Space Week and World AI Week!! So what better time to write a dedicated Space to Explore article on Artificial Intelligence for Space Exploration! This article aims to provide a broad analysis of the possibilities for using AI to augment space exploration and enable our long-term goal of humans on the surface of Mars.
As previous Space to Explore articles have analysed, the space exploration mission architecture guiding ESA, NASA, and other space agencies and actors plans to put humans on the surface of Mars before 2040. This requires moving exploration activities beyond Low Earth Orbit (LEO) and current operations on the International Space Station (ISS) to the orbit of the Moon (via the Gateway), whereby frequent missions to and from the surface of the Moon can be conducted. The Moon is a strategic stepping stone on the path to Mars, whereby new technologies for human and robotic exploration (including some of those in this article) can be developed and tested.
The graphic above provides an overview of upcoming activities and the overall mission architecture of the three major destinations (for more information on the architecture shown here, check out my previous Space to Explore article on The Future of Space Exploration). These ambitions for humans to reach Mars and travel beyond into our Solar System will only be made possible with the development and use of AI for Space Exploration.
A 101 on Artificial Intelligence
AI provides a series of techniques that enable systems to mimic some form of intelligence to complete tasks, such as performing data analysis or driving cars. Through AI, machines can respond to the sensory data they gather and make decisions on their own, allowing them to adapt to their environments and situations. This has a major potential benefit in that these signals do not have to be anticipated, and the systems don’t have to be controlled, by humans.
AI is often confused with any technology that performs tasks faster than the human brain in a robust manner. A calculator can perform complex mathematical equations at speed, but does that make it intelligent? Not in this context. Intelligence can be thought of as the ability of machines (agents) to perceive and recognise their environment, to be able to learn from past experiences, and logically make decisions based on new scenarios (usually whilst working towards a goal). Much like human intelligence, Artificial Intelligence consists of a spectrum of capabilities: a scale of speed, autonomy, and ability to generalise well to unseen situations that combine to create performance.
The fastest growing branch of AI is machine learning (ML) whereby AI models learn by themselves, in essence by training a relatively simple algorithm to become increasingly complex. A variety of ‘unstructured’ data forms, including images, spoken language, and the internet are used to carry out a range of activities from informing medical diagnoses, to creating recommender systems for the entertainment industry; from making investment decisions for financial companies, to helping driverless cars read highway signs. ML models process information in a similar way to humans by developing artificial neural networks (ANN or simply NN). The system progressively improves its performance on a specific task by “learning” from its environment, without being explicitly programmed.
Much like human intelligence, Artificial Intelligence consists of a spectrum of capabilities: a scale of speed, autonomy, and ability to generalise well to unseen situations that combine to create performance.
Within machine learning, a promising and advanced technique is maturing known as deep learning (DL), whereby the AI model, using multi-layered artificial neural networks, is able to train itself to perform complex tasks, such as image recognition. There are two main classifications of DL algorithms: supervised learning, where, for example, you feed the model pictures of celestial bodies until it can identify different types, or unsupervised learning where, the network finds the structure and writes the rules by itself, for example navigation systems for self-flying spacecraft.
AI can also be categorised by the generic function of techniques being deployed. It is possible to categorise AI into four major areas: (1) a foundational layer which encompasses traditional methods such as statistics and econometrics, complexity theory, and game theory; (2) a behavioral layer which serves as the operational processes such as process automation, machine translation, and collaborative and adaptive systems; (3) a sensory layer which provides information to the model in the form of language, audio, and vision; (4) a cognitive layer which provides the “intelligence”, including machine learning (deep learning), reasoning, and knowledge representation. These categories are good ways to consider the function of the type of techniques being deployed. The most sophisticated AI systems will use a combination of these.
The new dawn of AI
Artificial Intelligence is not a new phenomena. The field has experienced many waves of optimism and progress since its early form more than 60 years ago. Whilst many false dawns have been witnessed over the years, the progress of AI, particularly in the area of machine learning, of the past decade cannot be ignored. The recent resurgence of AI can be attributed to four main factors: (1) the emergence of data and the data economy; (2) sufficient increases in the computational power of technology; (3) the rise of new, more powerful, algorithms such as the DL reinforcement learning techniques aforementioned; (4) cloud computing and storage which has enabled easy access to an infrastructure of software and data that can be accessed globally.
Whilst its origins lay routed in academia, the rise of data science and the data economy has provided a demand for AI to significantly augment data collection, analysis, and most importantly decision making, and companies are starting to realise the value of AI as they strive to move AI out of the lab and into consumer products and services. AI is also starting to have a bigger impact on people’s lives, as it grows in society, and this impact only stems to grow in the near future.
The relationship people have today with technology will only develop further with AI, as their interactions become more personalised and goal orientated; with a greater dependency.
AI can be considered a suite of tools and techniques that can be applied to many different fields and used to solve a broad range of tasks. For instance, computer vision is having a huge impact on the entertainment industry, creating ever more realistic and adaptive video games. Deep learning has enabled fast and accurate speech-recognition and understanding on our phones. Natural language processing (NLP) and knowledge reasoning has enabled companies to create supercomputers, such as IBM’s Watson, that are able to analyse millions of text entries and convert hypotheses into actionable answers. Google’s use of NLP in its search engine is significantly increasing the time saved and value add of its service, helping customers find exactly the information they are looking for.
AI offers far more than just improving the technology in our phones, video games, and computers we use today however. In the coming years, AI will become a more prominent part of everyday life. The relationship people have today with technology will only develop further with AI, as their interactions become more personalised and goal orientated; with a greater dependency.
AI will help monitor the health and well-being of people, ensure their safety by alerting them to impending risks, and deliver services on a fast and “on-demand” level. Past fantasies, such as self-driving cars, are becoming a reality. Autonomous transportation will soon become widespread and will represent a significant breakthrough in public awareness and experience of AI, though the extent to which self-driving cars can operate is open to impending regulation. AI powered healthcare in the form of doctor assisted recommendations is ready to become a huge breakthrough in the medical industry, with AI models able to make very accurate and reliable diagnoses through pattern recognition and NLP techniques (though patient consent, data privacy and data security all represent major blocks in the use of AI in the public health system).
The use of AI and robotics is a powerful dynamic that is explored in this article for space exploration-and other uses, such as home/service robotics offer great potential to assist in a range of tasks, including elderly assistance, cleaning, entertainment, and security. AI robotics has the potential to create a new labour force in low-value industries (agriculture and factories) or harsh/dangerous environments (oil rigs, mines). These potential applications offer interesting parallels for this article and provide a good impetus for the synergies that exists across industries, including the growing sector of space exploration.
AI for Space Exploration
AI will have a significant impact that touches many different aspects of human and robotic exploration missions. The following graphic provides an overview of a number of different ways AI will augment our exploration activities as we progress along the architecture described above. This article will proceed to provide details about these.
AI Robotics
There has been a great history of robotic space exploration, with robots being used more frequently and able to do more every passing decade. Robots are used for a variety of different space applications, and their function can be divided into four main categories: Rover, Arm, Sampler, or Drill. The target for past and present robots has been for orbital applications, such as robotics in Low Earth Orbit (e.g. the International Space Station), and robots for planetary applications, such as robots on planets (e.g. Mars) or small bodies such as comets and asteroids (e.g. comet 67P/Churyumov-Gerasimenko-the successful destination of ESA’s Rosetta mission).
Robots have been used for space missions for decades due to their ability to work in harsh and hazardous environments that would be unsafe for humans. Robots handle tedious jobs very well and therefore in the case of human and robotic cooperation, can free up time for astronauts to work on higher value activities that require more complex problem solving and task orientation.
Robots are used for a variety of different space applications, and their function can be divided into four main categories: Rover, Arm, Sampler, or Drill.
AI builds upon this deep history and current developments in space robotics to offer promising augmentation capabilities to achieve complete autonomy, allowing for greater perception (vision) and dexterity which enables robots to make their own decisions. A variety of factors are driving space organisations towards increased automation for space. This includes the financial cost of human operation of space missions, a limited availability of operators and crew, the increasing complexity of systems and instruments deployed, and the many communication constraints, including limited bandwidth, windows, and latencies. With AI, the level of automation can extend far beyond the operational automation used in robotics over the past decades and can considerably augment the future of human and robotic spaceflight.
Whilst robots have been used for space exploration missions since 1967, the history of Artificial Intelligence has far later beginnings dating back to 1998 with the use of an AI algorithm called Remote Agent, used onboard Deep Space 1-a comet probe. Remote Agent, whilst elementary in comparison to the AI capabilities of today, proved its worth with capabilities including the planning and scheduling of activities and diagnosing onboard failures. Since then there have been many other applications of AI for space exploration missions ranging from algorithms that give greater autonomy to planetary surface rovers, AI for discovering new Exoplanets, and AI powered assistants aboard the International Space Station, which will be covered in the coming sections.
AI builds upon current developments in space robotics to offer promising augmentation capabilities to achieve complete autonomy, allowing for greater perception (vision) and dexterity which enables robots to make their own decisions.
AI will vastly transform robotics from its existing state of operating within known and static environments, to a new “intelligent” breed, able to move fast, with greater autonomy and an ability to generalise to unforeseen environments and circumstances. Advanced manipulation is an important benchmark of space robotics in order to carry out emerging space exploration applications such as resource utilisation and infrastructure building.
Currently, robotic navigation, particularly within known and static environments, is well matured. However in order to make these robots “intelligent”: able to move fast, with autonomy and generalisation to unforeseen environments and circumstances, AI offers much potential. In order to be intelligent, robots need to be able to interact with the environment around it. As you will read in the following sections, manipulation is an advanced form of locomotion that future robots must master in order to carry out emerging space exploration applications such as resource mining and infrastructure building.
The use of deep learning for robotics is just starting to emerge. AI models do not generalise well, and data on the geometry of lunar and martian surfaces is limited-therefore training these models is more difficult than a self-driving car which can test the same route a thousand times over. However, reinforcement learning (an emerging field within deep learning) offers an alternative to traditional hard-code and test models, whereby the AI model can learn to learn through experiences and reward system incentive structures-much the way a young child learns and develops. AlphaGo 2.0, the Go conquering AI system developed by Google DeepMind, was able to learn the rules of Go (the ancient Chinese board game) and beat the highest ranked champion through applying reinforcement learning techniques.
Reinforcement learning offers an alternative to traditional hard-code and test models, whereby the AI model can learn to learn through experiences and reward system incentive structures-much the way a young child learns and develops.
With a single objective function, such as to score the highest amount of points in the Go example, or traverse 100m across open terrain to point B in a space exploration context, an AI powered machine, using reinforcement learning, can become competent through its only learning and adaptation. This negates the constraint of having to “label” training data which is a requirement in supervised machine learning-i.e. it is not necessary to know and categorise (label) every variable (feature) in order for the model to be successful. This would clearly represent a huge breakthrough in environments whereby labelling features is difficult due to unseen variables. It is likely that a balance will be needed to achieve right trade-off between potential performance (full reinforcement learning) and reliability, safety, and robustness (supervised machine learning or pre-programmed parameters). Lunar and Martian like demo environments, such as the facility at Airbus in Stevenage UK, present an opportunity to test and train AI robots with a combination of supervised and unsupervised learning-tagging some features, but leaving scope for self-learning for adaptation to unforeseen scenarios.
Advances in machine perception, such as the use of more sophisticated computer vision algorithms will improve the success of robots in surface exploration, allowing them to have better vision and perception, which offers the potential for human-like interpretation of the surface. This would be a considerable improvement on classical robotic vision models, where perception consists of converting the input perception data into a geometric model before acting, which significantly slows down the transition between one action to the next. Computer vision models allow robots to “build up models of structures relating to perceptual domain to the agent’s actions”, therefore actions of the robotic agent can occur before perception analysis (scene representation). This enhanced perception is vitally important in instances where environments are unknown, allowing for robots to adapt to unforeseen scenarios. This improved vision, combined with force, and tactile perception will enable a new breed of AI robots that can achieve far more than existing exploration rovers and begin the robot revolution that enable large proposed projects, such as a lunar base (or settlement) to be realised.
There are two key areas where AI can significantly augment robotic space exploration: increasing autonomy and diversifying mobility (locomotion).
Increasing Autonomy
The level of autonomy of a robot is important because it impacts the range of tasks it can perform and thereby determines the demand and frequency of human robotic interaction. There are several different levels of autonomy on a scale, ranging from teleoperation to full-autonomy. Beer et al. 2014 provides a framework for categorising the level of autonomy which creates general guidelines for selecting and optimizing for the correct level of human robotic interaction, based on the purpose the robot needs to fulfil. In all but the final level, a degree of human intervention is needed.
For the purpose of this report, four generalised categorisation will be referred to regarding the level of autonomy: 1. Teleoperation (controlled from Earth), 2. Automatic Operation (pre-programmed controls), 3. Semi-autonomous Operation (start with predefined control sequence, where robot adapts to environment), and 4. Fully Autonomous Operation (autonomous decision making (goal orientated) e.g. distribute resources smartly etc). The figure from Gao et al., 2018 below presents an evolutionary roadmap of the capabilities of space robotics, with an increasing level of autonomy over time.
Diversifying Mobility (Locomotion)
Not only are robots becoming more autonomous, they are also becoming more mobile and dexterous, allowing them to carry out more complex tasks. For further exploration of the Solar System, such as visiting Europa, a Moon of Jupiter believed to have potential for forms of life, new advanced forms of robots, such as submarines, will be required to explore worlds with no hard surface in which to land. NASA are currently developing a helicopter called The Mars Helicopter, which is a rotorcraft to be deployed as part of their Mars 2020 rover mission.
To be able to use robots to support large scale programs in space, such as the construction of a lunar base, the ability of robots to manipulate the environment around them is a crucial step to achieve and requires advances in mobility, for example in areas such as drilling, lifting, and grappling.
In Space Navigation
Artificial Intelligence offers huge potential to augment in-space navigation, primarily probes exploring the extremities of our Solar System. Ground operation communication is too infrequent to allow for continuous human monitoring of changing and potentially hazardous situations. For a spacecraft to travel the vast distances of outer space (interstellar travel), greater autonomy is a huge enabler for increased mission complexity and chance of success. Human intervention is difficult and limited given the huge lag in transmission across space. During interstellar missions, probes cannot wait for mission control on Earth to instruct them. Therefore an AI command computer onboard the probe is necessary to operate the probe in navigation to deal with course corrections and communications. More sophisticated future AI powered probes would be able to fulfil more complex operations such as dealing with faults and other unplanned incidents, conducting maintenance and upgrades over the longevity of the life of the probe.
Human intervention is difficult and limited given the huge lag in transmission across space. During interstellar missions, probes cannot wait for mission control on Earth to instruct them.
Hein and Baxter (2018) also identify instances whereby complex navigational calculations can be made in real-time by the probe to react to conditions difficult to predict from Earth pre-mission, for example in the case of probes such as Icarus and Dragonfly visiting Alpha Centauri (the nearest star beyond our Solar System), due to the nature of “coming in from out of the plane of a double-star system, [whereby] a complex orbital insertion sequence would be needed”, accompanied by the deployment of subprobes and the coordination of communication with Earth.
Teaching a spacecraft to make a controlled landing without human input is an important step in space exploration ambitions to send rovers to explore distant worlds such as Jupiter, where the lag between Earth and spacecraft communications is too great to allow for a manual human operated landing. Closer to home, autonomous landing, rendezvous, and docking capabilities, such as docking a small spacecraft with the ISS, would open up a fleet of spacecraft and satellites that could carry out their function to explore, transmit, and observe, and then perhaps even return to a station for maintenance and upgrades. Advanced navigation, such as autonomous rendezvous and docking requires improved guidance, navigation and control (GNC) algorithms, along with improved docking and capture mechanisms and interfaces for multi-agent coordination. Reconfigurable and adjustable autonomy is a crucial element in enabling this.
Advanced navigation, such as autonomous rendezvous and docking requires improved guidance, navigation and control (GNC) algorithms, along with improved docking and capture mechanisms and interfaces for multi-agent coordination.
Artificial Intelligence enables local decision-making which sets future “intelligent” probes apart from their counterparts. The key benefit to emphasise is an ability to handle unforeseen and unpredictable scenarios with some autonomy onboard the probe. Even rather basic intelligence for making minor corrections, such as countermeasures to respond to failures, e.g., a failure to deploy solar panels, could be the difference between mission success and failure. Another area that could be useful for future space exploration is automated collision avoidance manoeuvres. Machine learning (ML) algorithms can identify debris in space so that decisions on collision avoidance can be anticipated and avoided.
Planetary Navigation
Robots can self-navigate and thrive in environments too dangerous or not possible for humans, such as planetary surfaces, using laser scanners and wheels that turn into legs to navigate. Rovers can navigate around obstacles independently of human control and intervention, even across unknown fields, using advanced forms of computer vision. On Mars, the ability for robots to drive autonomously is a technology that has existed for over a decade. An autonomous navigation system called AutoNav has been used on previous rovers-Spirit and Opportunity, and is part of the navigation and driving system onboard Curiosity.
Perception requires building understanding of the environment based on the sensor inputs to provide situational awareness for space robotic agents, explorers and assistants. However current sensing relies on steer camera (and some instances of Lidar) which is very slow and takes a lot of computation for 3D reconstruction. New sensors and sensing techniques are needed, such as algorithms for 3D perception, state estimation and data fusion. Data processing is a significant challenge currently, due to the huge volume of data that needs to proceed and therefore better onboard processing systems will be required that reduce the energy and storage requirements for this data processing. Better computer vision algorithms will help tackle object, event and activity challenges relating to responding to “unseen” events that occur outside of the model training.
Data processing is a significant challenge currently, due to the huge volume of data that needs to proceed.
Future RoboExplorer’s will need to be able to do more than just self-navigate and drive, in order to fulfil requirements for future space exploration missions that require: surveying, observation, extraction of resources, and deploying infrastructure for human arrival and habitation. The following section on In-Situ Resource Utilisation explores this further.
Future AI robots will need to improve beyond basic rover movements to incorporate new forms of mobility, including: walking, flying, climbing, rappelling, tunnelling, swimming, and sailing.
In-Situ Resource Utilisation
The use of AI and robotics for the mining industry on Earth serves as a good analogue for in-situ resource utilisation (ISRU) on the Moon, Mars, and beyond. The implementation of AI and related autonomous technologies for mining on Earth started around a decade ago with the introduction of autonomous trucks. In-Situ Resource Utilisation can be divided into two key areas: (1) prospecting and exploration; (2) mining. AI can have a significant impact on both.
AI models can be fed the geological, topographical, mineralogical, and mapping data to be able to automate the prospecting process to find areas of interest
Prospecting is the process of searching for identifying the specific resources you want to utilise, and on Earth requires an economic assessment to determine whether the investment is viable or not. Exploration involves sampling and analysing the prospect to determine its value. Both of these stages require collecting vast amounts of data which, done manually, can take years to analyse. AI models can be fed the geological, topographical, mineralogical, and mapping data to be able to automate the prospecting process to find areas of interest, and increase the level of exploratory certainty to send robots to begin the sampling process.
AI can augment to second major process of mining through the development of autonomous drills on autonomously navigational robots that can easily perform drilling activities faster and more efficiently than the currently manual process that exists on Earth today. For large scale mining operations, such as a future lunar base, where humans and robots will work in tandem, AI robots can lift the burden of the tedious, hazardous, and heavy duty tasks away from humans who can focus on more important, value adding tasks that require greater intelligence, such as scientific analysis. AI robotics used for mining can also benefit from sensing technologies that allow the machines to monitor the atmosphere and hazards, and send warning signals, having detected anomalies and dangers that aren’t self-evident to humans.
Future AI robots will need to improve beyond basic rover movements to incorporate new forms of mobility, including: walking, flying, climbing, rappelling, tunnelling, swimming, and sailing. These new forms of space robotic mobility would dramatically increase the ability of planetary robots to carry out more extensive activities, contributing to new fields such as in-situ resource utilisation, infrastructure building and development, and long distance surface exploration. This will require manipulations to the environment and of objects using motions such as digging, drilling, sampling, grappling, placing, and assembling. These new forms of mobility will require more sophisticated forms of perception, vision, and movement.
AI robots can lift the burden of the tedious, hazardous, and heavy duty tasks away from humans who can focus on more important, value adding tasks that require greater intelligence, such as scientific analysis.
Communication
Prioritising communications for probes on long-duration missions, including efforts to send probes on interstellar missions is a crucial function enabled by AI powered probes. Current communications with the ISS experience a very small time delay-less than a second for radio waves to travel. However, for probes exploring the far regions of the Solar System, this delay is far greater. For the speed of light alone-the fastest known particles in the universe-the time delay is between 6–42 minutes to travel to Mars and back and over an hour beyond Jupiter. Given this lag, it is not practical for Earth to relay communications.
The recent development of cognitive technologies is an advancement in the architecture of future communications systems. “By applying artificial intelligence and machine learning, satellites control these systems seamlessly, making real-time decisions without awaiting instruction”, Janette C. Briones, principal investigator in the cognitive communication project at NASA’s Glenn Research Center. AI can enhance communication networks by picking out “white noise” in communications bands to transit data which maximises the use of limited telecommunications bands available, and minimises delays. AI can therefore find underutilised portions of the electromagnetic spectrum without human intervention. “We envision these technologies will make our communications networks more efficient and resilient for missions exploring the depths of space. By integrating artificial intelligence and cognitive radios into our networks, we will increase the efficiency, autonomy and reliability of space communications systems.”, Janette C. Briones.
For the speed of light alone (the fastest known particles in the universe) the time delay is between 6–42 minutes to travel to Mars and back and over an hour beyond Jupiter. Given this lag, it is not practical for Earth to relay communications.
RoboSats have been proposed for space exploration missions. This includes the use of swarms of small robots that benefit from AI algorithms for communicating amongst one another in the form of a “network”. RoboSats can learn iteratively through trial and error to take actions that lead to high overall system return, rather than being programmed what to do (multi-robot coordination through reinforcement learning). This is also known as hive learning. This leads to improve satellite coordination and operations as fleets of small RoboSat constellations, bring flexible to operations, including relative positioning, communication, and end-of-life management, e.g. providing a global navigation system by flying in formation. It is also possible to do more which such as fleet of satellites, such as conducting more complex navigation and positioning, for instance, blocking the Sun with one satellite to observe its corona with another.
Developing cognitive communications is important to increase the efficiency, autonomy, and reliability of communication architecture. In the table below, four major categories for cognitive augmentation are set out for the NASA Space Communications and Navigation (NASA SCaN) communication architecture. The key challenge that AI can help overcome is to autonomously optimise the links for maximum data throughput-creating autonomous networks in the process.
For use of an ultraviolet imaging spectrograph (UVIS) on Europa, one of Jupiter’s largest moons with a water ice surface, future missions by NASA intend to scan Europa for fresh plumes to detect sources of life. The important part is they to target observation on that flyby without the ground in the loop. Why? Because If a plume isn’t caught in that instance and the spacecraft has to wait until the next time it comes around (for Europa this could be around 30 days), it’s more than likely the complex organic materials may have already broken down from the harsh radiation environment. Therefore autonomy without Earth communication is imperative for the chance of detecting life on Europa and conducting science in the far extremities of our Solar System.
Mission planning
One of the most effective uses of AI for space exploration occurs at one of the least exotic applications: the business of mission planning and scheduling. Spacecraft, for instance a Martian rover, require extremely complex scheduling systems, with every detail, action, and mechanic meticulously planned. This task requires whole teams of people who work many hours to fulfil this need. As outlined above, the ability to give robots objective functions that they can intelligently execute, such as explore crater “x” (target) at “y” (location), negates the need for such fine planning. The rover’s onboard AI powered computer can plan the steps and reroute itself based on its experiences of the evolving environment.
A Martian rover, require extremely complex scheduling systems, with every detail, action, and mechanic meticulously planned.
AI can be used as an operations tool for space missions and for planning missions for deep space probes, as algorithms can find optimal trajectories. With the optimal trajectory, AI can be used to create an automated planning schedule, by solving a series of multi-objective optimisation problems. The model can deal with key scientific and engineering missions constraints and optimise the schedule to best meet the needs of all involved. Such a system was used in the medium term phase (MTP) of mission planning for the ESA Rosetta mission, specifically in the area of SK-where the trajectory is known and scientific measurement campaigns can be scheduled. AI is used to rapidly iterate the alternate measurement strategies, which feeds into a manual planning tool (called MAPS at NASA). The manual decision making is taken at the sequence level which are the low level commands that actually get communicated to the spacecraft. Manual planning and scheduling is difficult. For the Rosetta mission, over 2000 observations (different slews of the craft) were calculated. In addition there are many irregularities of space. Geometry of the comet was a big challenge on the Rosetta mission, as the bi-lobed shape of the comet wasn’t known until the Rosetta spacecraft caught up to it. Hundreds of different vectors and geometries make planning observations-for hundreds of scientists across multiple agencies and countries a real challenge.
AI can play a key role in conducting risk analysis assessments to quantify the level of mission and operational risk inherent.
To complicate matters further, ground communication network uplinks and downlinks have to be integrated into plans. NASA’s Deep Space Network (DSN) provides communications for planetary exploration missions. However, the process of scheduling the DSN is highly time-consuming and complex. There is significant demand for the DSN beyond what is currently handled by the available assets and therefore AI presents a potential solution to this optimisation problem. AI can help with the automated scheduling of activities.
The other area where AI can play a key role is conducting risk analysis assessments to quantify the level of mission and operational risk inherent. Predictive models can be created with AI which allow space actors to more effectively assess, prioritise, and manage their risk.
Mission operations
In additional to mission planning and scheduling, AI can also augment the mission specific operations in a number of ways. AI can help perform spacecraft monitoring to assess the state of the spacecraft (e.g. probe) and ensure that all subsystems are performing as expected. The real value is in automating these checks and allowing the AI to find anomalies in the data that may be missed by humans. Managing the power subsystem of a spacecraft is a key area where AI can support human controllers, in particular with off-nominal situations.
The real value is in automating mission operation checks and allowing the AI to find anomalies in the data that may be missed by humans.
RoboAssistants have a long history in sci-fi films, with human and machine operating together. Supercomputer systems, such as IBM’s Watson, have been developed with applications such as medical support for doctors in mind, to help augment the doctors ability to carry out their job. With CIMON, this is now in the early days of reality. In June 2018, CIMON, which stands for Crew Interactive Mobile Companion, was launched to the ISS. CIMON is an IBM technology, building upon Watson, and developed in partnership with the German Aerospace Agency (DLR) and Airbus, which aims to augment the information and learning available to astronauts aboard the ISS. CIMON’s main function is to present tutorials on how to do things. For instance, in the case of conducting an experiment, CIMON can aid astronauts by answering questions such as: what kind of tool do I need to use? Or why must I use Teflon over another material? The main focus is on how RoboAssistants can help astronauts perform experiments, e.g. CIMON supporting crystallisation experiments, and guide them through the procedure.
CIMON aims to augment the information and learning available to astronauts aboard the ISS.
As well as helping with experiments and procedures, AI can also be a very powerful education tool that can be used to create bespoke training for astronauts. Every hour of time spent onboard the ISS is estimated to cost €100 million per ESA astronaut, and therefore training (or retraining) is very rarely carried out onboard the station. However, after a fast and thoroughly comprehensive three years of training (consisting of basic and mission specific training), astronauts would benefit from refresher training on board the station. AI can provide personalised training (through NLP, Bayesian Knowledge Tracing, and ML), specific to the needs of a particular astronaut. If connected to other systems on board the station, the AI can help educate and support astronauts as they carry out tasks-providing an environment for training and learning which doesn’t significantly disrupt the busy schedule of tasks the astronauts are working towards.
In the future, day-to-day operations on planetary surfaces, for instance in the case of a lunar settlement, is another area where AI can help. AI algorithms can be deployed to monitor for risks and dangers and proactively alert relevant parties.
Robonauts
NASA defines a Robonaut as “a robotic astronaut assistant”. A Robonaut is designed in a structurally anthropomorphic manner, with the target dexterity, manipulation, and size as an astronaut. Using Robonauts over other forms of robots to conduct similar tasks has the distinct advantage of complying with the standardisation architecture and operations for space, such as the ISS, which is designed to accommodate astronauts.
One of the most hazardous activities carried out by current astronauts aboard the ISS is a spacewalk-a venture outside of the station, usually to install or repair something on the exterior. Along the same rationale as using autonomous robots for exploring hazardous planetary environments, Robonauts can be deployed to carry out these tasks, significantly reducing the current risk to astronaut safety today. Spacewalks are a timely and costly endeavour, with astronauts required to prepare extensively before leaving the station-a routine which includes pre-breathing at space suit air pressure for up to 4 hours.
Robonauts can be deployed to carry out dangerous tasks, such as spacewalks, significantly reducing the current risk to astronaut safety today.
Robonauts have already been designed and sent aboard the ISS, such as the NASA Robonaut 2 in the photo below. Whilst this might suggest the technology is ready today, the technology is very primitive and lacks true intelligence. Therefore, the use of Robonauts for their intended function is seen as a far-term application that AI can support, through a range of computer vision and perception, reasoning, and advanced manipulation techniques.
Data collection and analysis
One of the most powerful applications for AI to augment space exploration exists in the process of collecting and analysing the huge amount of data gathered from prospecting, sampling and scientific discoveries.
AI can help detect and characterise interesting features such as usual and static, including water, ice, and snow, and unusual and dynamic, which includes features like fires, volcanic activity, floods, plumes, etc.
Automated data analysis for decision making offers huge potential to significantly augment the function of spacecraft in outer space. Probes would benefit from AI for data collection to optimise the points in time at which the probe should collect data, such as photos, on a particular mission. AI can help detect and characterise interesting features such as usual and static, including water, ice, and snow, and unusual and dynamic, which includes features like fires, volcanic activity, floods, plumes, etc. From this AI algorithms can autonomously create environmental maps. AI can even be used to analyse the data gathered onboard the spacecraft to determine the most useful and significant data to transmit back to Earth. This onboard analysis is important to assess which data collected is desirable and can therefore save storage by discarding data that has little or no scientific relevance.
On Planetary surfaces, AI algorithms can help with sampling and in conducting scientific experiments. AI is already being used to power the Curiosity rover on Mars, through an algorithm called AEGIS (Autonomous Exploration for Gathering Increased Science), which enables Curiosity to identify rocks and terrain features of scientific interest and relevance, autonomously. Due to the low data transmission rates and power constraints, rovers cannot send back to Earth all photos taken, and therefore by intelligently choosing the most valuable targets through AEGIS, the rover can maximise the quality of scientific data collected. The use of deep learning algorithms for pattern recognition and object classification for on-board processing can further the extent of scientific and geological analysis performed in the coming years. As the rovers become more intelligent, they can process raw sensor-data into scientific knowledge and actionable insights.
AI can even be used to analyse the data gathered onboard the spacecraft to determine the most useful and significant data to transmit back to Earth.
AI is also already being used by organisations such as Planet to analyse the data on the ground that they receive from their fleet of Earth observation satellites. Machine learning can be used to sift through and analyse Earth observation data to extract meaningful information that furthers our understanding. For instance, analysing observations from the Sentinel satellites with machine learning algorithms and sifting through data from ESA’s Gaia satellite to find out more about stars. AI has already had great success in analysing satellite data to find and classify exoplanets.
Knowledge discovery
Deep learning models can discover the environment for themselves, write their own rules and go off in different directions contrary to human expectations and what humans perceive to be “correct”. Machine learning often necessitates an “expert”, in many cases team of expert scientists, to create “feature extractors” which enable the model to learn. However deep learning models are able to find these features by itself which is a huge advantage in the area of scientific discovery where humans do not know what to look for and have incomplete information. For example, deep learning algorithms were used by NASA to discover two exoplanets previously missed by humans.
Deep learning models are able to find their own features of interest which is a huge advantage in the area of scientific discovery where humans do not know what to look for and have incomplete information.
AI models not taught by humans are free from human limitations and bias and often find surprising and creative results. In the future, this could lead to probes that are sent into deep space, powered by deep learning algorithms, that can perform autonomous science experiments, by collecting and analysing data, forming their own hypotheses and sending results back to Earth.
Spacecraft maintenance
If we are serious about living and working in space for the long haul, we are not going to discard our hardware every time it breaks down or runs out of propellant. When spacecraft fail, run out of fuel, or reach their end of mission, they are discarded and new spacecraft are developed. The ability for spacecraft to carry out maintenance, on themselves and other spacecraft would present a significant breakthrough in the longevity of space exploration missions. Maintenance includes everything from carrying out repairs, to refuelling, and installing upgrades.
If we are serious about living and working in space for the long haul, we are not going to discard our hardware every time it breaks down or runs out of propellant. The ability for spacecraft to carry out maintenance, on themselves and other spacecraft would present a significant breakthrough in the longevity of space exploration missions.
Intelligent robots are already being designed to inspect and service in-orbit satellites. Activities include refuelling, upgrading electronics, extending jammed telescopic antennae and unfolding solar panel arrays that failed to deploy. The same technology can be applied to space exploration, where a special breed of maintenance or “servicing” spacecraft can be deployed to autonomously dock and carry out works on another spacecraft. Future orbital servicing robots can also be deployed for assembling large space telescopes, manufacturing in space, and to assist in the removal of space junk.
For future deep space exploration missions, the use of servicing spacecraft would not be feasible. In this instance, a spacecrafts ability to conduct self-maintenance is paramount. For such missions, AI can be used to perform the problem analysis and diagnosis, and then intelligent robotics aboard the spacecraft would need to be able to fix the problem-such as the redeployment of failed solar panels.
Astronaut health monitoring
Outside of the space sector, one of the most promising emerging applications of AI is in the field of healthcare, where the use of advanced pattern recognition software for diagnosis and deep learning algorithms for scientific discovery, is having a big impact-enabling new insights to augment the role of doctors and health workers. AI-based medical care can also be applied to space exploration. As humans ready themselves to leave LEO, new medical care systems that make astronauts more autonomous for delivering their own care, are highly important.
Monitoring and assessing the health of astronauts on long-duration missions is a crucial challenge for space actors currently. Health is a huge barrier to overcome on deep space exploration missions to Mars. The growing use of wearables to monitor biometric data of users and analyse indicators, such as heart rate and VO2 max has become customary for many people worldwide. Health and fitness wearables is a huge market. AI offers huge potential to augment the value gleaned from such monitoring, by making assessments and recommendations based on the health data gathered. AI can be used to identify and classify health data in real-time to support the efforts of on the ground medical officers.
As humans ready themselves to leave LEO, new medical care systems that make astronauts more autonomous for delivering their own care, are highly important.
The monitoring of biometric data can work in tandem with computers such as CIMON that can deploy NLP techniques to analyse health data and potential symptoms against global health records to make a diagnosis. Whilst this would not negate the need for on the ground medical expertise, the ability to make early warning predictions, before symptoms grow and develop would allow for quicker preventative measures and would represent a significant breakthrough in ensuring the health and safety of astronauts on long-duration missions. AI can also vastly improve the ability to conduct accurate predictive analytics that assess long-term trends and predict future health problems. This can lead to personalised health plans (exercise and diet) that adapt to the health of the astronauts over the duration of the mission-proposing what to do, when to do it, and how frequently it should be done which can feed into mission operations.
AI can vastly improve the ability to conduct accurate predictive analytics that assess long-term trends and predict future health problems.
For deep space exploration missions, AI robotics could be extended to surgical robots designed to assist medical procedures. Given the nature of the broad capabilities that astronauts are required to possess, astronauts are only trained to carry out very basic medical procedures in space (using the ISS as a point of reference). With this in mind, many missions are designed to ensure a medical doctor is part of the crew. The ISS is equipped to conduct only very basic procedures, and due to the microgravity environment, any complex procedure would carry a significant risk of operation and recovery. AI powered surgical robots could carry out some autonomous procedures which, given the high degree of accuracy, could reduce some of the operational risk, and provide solutions for carrying out medical procedures in instances whereby a medical doctor is not available.
Design and Construction of Artifacts
One of the major future applications of AI robotics is the ability of machines to create artifacts such as spacecraft, and infrastructure to enable the creation of bases or settlements on other celestial bodies. The construction of bases will enable further space exploration by creating centralised capabilities in space from which further missions can embark. A number of space agencies are looking into building a base or settlement on the Moon to support cislunar and deep space activities.
Currently, the sophistication of robot locomotion, including manipulation, and intelligence is not sufficient to enable robots to be able to create spacecraft fully autonomously, in the same way robots play a huge role in the production of cars in the automobile industry. And whilst many plans to build a lunar base consist of a fully robotic phase to prepare the infrastructure for human arrival, the ability of robots to construct complex compounds is far from reality. However, other technologies such as large industrial 3D printers offer potential to construct modules in a more efficient manner-the ability to print basic houses has already been developed on Earth.
About half the energy required to get anywhere in the solar system is needed to leave LEO. There is an abundance of structural metals found on the lunar surface (e.g., titanium, iron, and aluminium), which could be used for in-space manufacturing. The manufacture of spacecraft structures in space would allow for lighter and cheaper rockets, as rockets currently made on Earth require stronger materials to withstand launch and gravitational forces. In the future it could be possible that robots can build other spacecraft completely autonomously on Earth but crucially also in space. Building on space means new types of spacecraft could be manufactured and launched from a base. For instance, structures could be much bigger and carry a larger payload than those we see today. With the surge of nano satellites, these new spacecraft can be deployed from celestial bodies could transport huge constellations of satellites manufactured in space for deep space missions.
Self-assembling Spacecraft
Building on from robots that can build other spacecraft, this application considers the potential for future spacecraft and robots to be able to self-modify their structure to achieve a wider range of tasks. Currently, AI is generally considered to have “narrow” applications, in other words, an AI model is designed and trained to carry out a specific task. In the future, the expectation is that AI will advance towards a concept referred to as AGI-Artificial General Intelligence, whereby the machines become far more intelligent with a range of capabilities and skills-much like humans. They won’t be specifically designed and trained to carry out individual tasks, but rather have a general level of intelligence that can be applied to solve many tasks.
The ability of robots and spacecraft to self-assemble and self-modify has huge potential for deep space exploration missions whereby a range of locomotion and manipulation is required to carry out a broad set of scientific missions. This would mean that rather than designing a robot for a particular function, for instance a drilling robot, or for a particular terrain, e.g. submarine, the robot/spacecraft can restructure (self-assemble) itself into new forms to carry out different tasks. And this would only be possible with intelligence in the form of AI algorithms, sensing the environment and instructing the robot to make the changes, and then having the ability to ensure the changes have been implemented correctly and the robot/spacecraft functions correctly. This would require sophisticated AI models with advanced simulation, optimisation and reasoning capabilities.
Self-Replicating Spacecraft
A self-replicating spacecraft is designed to be able to create copies of itself by using the materials and energy in space. The idea of a self-replicating machine, though formally proven to be possible, is very difficult to achieve in practice and has long been held as science fiction fantasy. However, with the emergence of AI, this fantasy may be a step closer to reality. In the world of AI, it has long been thought that the most powerful creator of AI models and systems will be AI itself. As per the discussion on Knowledge Discovery, AI (specifically deep learning) has a distinct advantage over humans in that it can compute millions of complex functions at once, free from human limitations and biases of what we expect to observe. Humans generally aren’t good at thinking outside of normality, and have a huge bias for previous.
The idea of future AI systems creating AI systems can be extended to self-replicating spacecraft, where spacecraft can create copies of themselves, “descendants”, to spread around the galaxy. This offers a potentially affordable way of exploring vast regions of space but also a sustainable way, whereby once a spacecraft reaches its destination, it self-replicates-creating a chain of spacecraft.
Communication with Extraterrestrials
Building on from self-replicating spacecraft, AI has long been proposed as a medium for contacting and communicating with extraterrestrials. This idea dates back to 1960 when Bracewell imagined sending out a fleet of probes powered by AI (with human level intelligence) to explore and attempt to make contact with possible extraterrestrial races.
The AI would present an alternative to the conventional SETI (Search for Extraterrestrial Intelligence) method of detecting electromagnetic signals. Bracewell proposed that an AI powered probe could initiate contact and send back native signal from the extraterrestrial being using radio technology. However, modern day AI may be capable of far greater communication, made possible by huge strides in recent years in the areas of deep questioning and answering, machine translation, and even audio and speech. The AI would be able to not only send a signal, but also potentially interpret a signal or action of the extraterrestrial agent.
Challenges to Overcome
AI Readiness
The ability to optimise the function of a robot or AI system depends primarily on the Technology Readiness Level at the time of deployment. Artificial Intelligence is an evolving field, whereby new forms of deep learning are showing great promise, yet lack significant scalability and cannot be relied upon currently. The ability to achieve some of the emerging and future applications of AI for space exploration outlined in this article depend upon the AI models and methods to keep improving in order to realise the potential.
Due to the complex nature of space exploration, there exists a lag between technology readiness on Earth, in typical industries such as retail and food where AI can be more quickly implemented and tested, such as marketing, demand modelling, and supply chain logistics, and for space. Emerging industries, such as the market for self-driving (autonomous) vehicles, offer good testing grounds for AI computer vision and perception models which can be improved and optimised for use in outer space. The lag in the application of the technology between Earth and Space in the area of autonomy exists for several reasons, including but not limited to: a lack of opportunity to optimise deep learning algorithms on training data of planetary applications (an important part of the ML/DL process), a lack of power and storage for data analysis and transfer, and a low level of global explainability to decipher why an AI robot performed a specific task-an important requirement when trying to optimise the robot to perform better; made complicated by an inability to directly observe a robots actions.
Whilst autonomous drones provides an example of the state of existing technology on Earth, there are significant limitations to deploying this technology for deep space exploration missions. Currently, the mechanics of robotics does not allow for such dexterous movements, however companies such as Boston Dynamics are working towards improving this.
Spacecraft Computing Power
The computational power on existing space probes is extremely limited, and hence they process information very slowly. To give an example, the computer onboard Curiosity is a CPU called RAD750 which has the computational performance of a computer from the late 1990s, with a speed of 200 Mhz and only 256 MB of memory (RAM). Compare this to a standard, modern day household computer with a speed of 2 to 3 GHz and a RAM of 8 to 16 GB. A huge difference.
One of the key limiting factors therefore to achieving many of the remote applications of AI outlined in this article (navigation, ISRU, data gathering and analysis) is computing power onboard spacecraft-something which AI algorithms need a lot of in order to work. One of the key limitations preventing the use of more powerful computers, aside from mass budget constraints is radiation. The development of radiation-tolerant CPUs take time to develop and test, and can be a costly endeavour.
Acceptability Challenges
It must be emphasised that whilst AI development has come a long way in recent years, the most valuable models, which are inherently complex, currently lack robustness and reliability.
Reliability and Robustness
AI is unproven in many aspects and lacks sufficient testing and proofing to be reliable. AI makes mistakes, and the more sophisticated the AI and the less supervised it is, the more likely it will make mistakes as it learns and develops. Space exploration missions are very expensive, and take years of proposals, planning and preparation. Therefore, one of the key obstacles to overcome is stakeholder trust that the AI can deliver on its promises. And with a high degree of risk pointing towards potential mission failure, it’s a risk that some organisations, particularly those funded by national budgets, may be unwilling to take.
AI models used for space exploration require a 100% robustness guarantee to satisfy project executives-for all remote applications the software simply cannot fail. For these reasons, performance is often traded off for additional robustness. A key enabler to achieve robust and reliable results is interpretability. However, greater model complexity (performance or in AI language “accuracy”) which is mainly determined by the type and size of AI algorithms deployed (e.g. large deep neural network vs a small decision tree) usually results in reduced interpretability (see figure below on the relative explainability of learning algorithms).
Due to the need for engineers to take on as low additional risk as possible, advanced forms of AI, such as deep learning, are on the whole considered to be too unreliable and lack the necessary robustness to guarantee mission success. Therefore even if the technology is available and ready and the craft or rover could support the requisite power mass, the high risk of failure may convince engineers to design and operate the machine as conservatively as possible, with human operators manually controlling it with constant checks and approvals.
It is likely that a balance will be needed to achieve right trade-off between potential performance (full reinforcement learning) and reliability, safety, and robustness (supervised machine learning or pre-programmed parameters).
This restricts the level of autonomy the machine can operate at. Currently, rovers used on Mars, such as Curiosity, have a high degree of inbuilt autonomy-theoretically they are able to traverse long distances unaided. However, the AI algorithms (AutoNav) are designed extremely conservatively in order for the operator to stay in near complete control of every movement. Upon the slightest detection of an anomaly, the rover stops in its tracks and activates a ‘safe’ mode whereby it lies in wait of human operator instruction. This is a slow and arduous process. In a typical day, a rover using AutoNav on Mars with these autonomy constraints will travel 10–20 metres per Sol (a Martian day of 24h 40m). However, this limitation exists in part due to the aforementioned reliability and robustness challenges. And as mentioned previously, due to the requirement for 100% robustness, given the nature of space missions, AI algorithms will be restrained and measured against a very high standard of reliability until they are trusted.
Model Explainability
As AI continues to grow in sophistication, complexity and autonomy, one of the major challenges is the ability for AI ML applications to operate without black box scenarios, whereby the model offers no discernible insight into how decisions were made and why a particular decision was chosen. This is a significant barrier to applications where the reasoning is crucial, such as knowledge discovery, or in instances where AI models are designed to learn from their experiences and environments overtime, such as reinforcement learning techniques deployed on planetary rovers, whereby an understanding of why the robot made a certain decision is fundamental in the training and development of the robots. Explainability is therefore an important factor in being able to improve performance-understanding why and how the AI works enables the designer to fine tune and optimise the model.
A key research area amongst the AI community, and a previous work from this author, is that of Explainable AI (or ‘XAI’). An ML application is considered “explainable” if it provides the user with a degree of qualitative and functional understanding, otherwise known as ‘human style interpretations’. There are two broad degrees of explainability which should be satisfied: (1) global explainability, which enables the user to understand how the input features (variables) affect the output of the model; (2) local explainability, which provides an explanation for why a specific decision was made-such as a particular movement of a planetary rover in the previous paragraph. XAI aims to make AI models more interpretable, bringing greater transparency to decision making-a crucial step to building trust in using AI for space exploration missions.
Conclusion
In conclusion, Artificial Intelligence is having a growing impact in many sectors, from monitoring and diagnosis in healthcare, to recommender systems in entertainment, and from self-driving technology in transportation, to fraud and error detection in financial services. There exists a continuum of AI capability from basic augmentation of data analytics and prediction, to intelligent process automation, and fully cognitive machines. Much of the recent progress in the field of AI can be attributed to a few major factors, including the availability of data, sufficient computing power, cloud-based computing architectures, and advances in machine learning algorithms.
With current technology deployed by space agencies, spacecraft must communicate with Earth to function. However, the introduction of intelligent autonomous spacecraft that can make their own decisions: for navigation, guidance and control, for communication, telemetry analysis, function (e.g. when to take an image), power, and emergency response, would drastically lower the risk of missions, and increase the value of space exploration. These spacecraft would in essence be able to look after themselves and make real-time decisions based on their environment and situation, something not possible through direct control on Earth due to the lack of visibility, the required amount of information, and communication time lag. Missions would deliver a higher return on investment at a reduced level of risk of failure.
One of the most evident uses for AI in space exploration is the need for intelligent exploring machines that can make decisions on their own in remote, potentially hostile environments. Systems with AI can perform tasks that usually require human intelligence. These automated operational activities can free up human time in space for more high value and rewarding activities.
It is still early days in the development of ML models and common examples used extensively for space applications are sparse, as the models developed are lack reliability and robustness, and are not globally explainable by design (meaning the overall model is not understandable to a human). The tools to interpret these models are in their infancy however strong academic focus is being applied to this area. Other factors including the processing speed and memory of spacecraft and the necessary level of robotic manipulation present considerable constraints to many of the applications outlined in this article.
AI has been used relatively modestly, given the high-risk of space exploration missions, and early applications in the industry have focused on supportive functions such as the data analysis of Earth observation and spacecraft telemetry data, where ML can play a big role in sifting through huge amounts of data, at a much faster and more accurate rate than humans. However, in order to achieve the current ambitions goals of space agencies worldwide, AI will be relied upon to help make the mission a reality. When going to Mars, spacecraft, robots, and astronauts will all rely on AI. Deep space missions to explore the Solar System, because of transmission and power supply constraints and time lag, will rely on AI for optimising the decisions of which data is sent back.
Artificial Intelligence is here and ready to fuel a new dawn of space exploration. AI enables space actors to do more; better; for less. AI is not just the future, it’s the present. And it’s ready if you are.
Thanks for reading this long article, I hope you found it interesting and useful. I’ll be at the World AI Summit over the next couple of days (October 9/10th) so please do come and find me! And be sure to look out for a host of exciting Space and AI related events this week.
— — — — —
Beer, J. Fisk, A. D, Rogers, A, 2014. Toward a Framework for Levels of Robot Autonomy in Human-Robot Interaction.
Briones, J. C, 2018. Artificial Intelligence — The Future of Space Communications.
Gao, Y, Jones, D, Ward, R, Allouis, E, Kisdi, A, 2018. Space Robotics & Autonomous Systems: Widening the horizon of space exploration.
Gunning, D, 2017. Explainable Artificial Intelligence.
Hein, A, Baxter, S, 2018. Artificial Intelligence for Interstellar Travel.
Hein, A.M., 2016. Artificial Intelligence Probes for Interstellar Exploration and Colonization.
