Enabling and Enhancing Astrophysical Observations with Autonomous Systems

Rashied Amini1,a, Steve Chien1, Lorraine Fesq1, Jeremy Frank2 , Ksenia Kolcio3, Bertrand Mennsesson1, Sara Seager4, Rachel Street5 July 10, 2019

Endorsements Patricia Beauchamp1, John Day1, Russell Genet6, Jason Glenn7, Ryan Mackey1, Marco Quadrelli1, Rebecca Ringuette8, Daniel Stern1, Tiago Vaquero1

1NASA Jet Propulsion Laboratory 2NASA Ames Research Center 3Okean Solutions 4 arXiv:2009.07361v1 [astro-ph.IM] 15 Sep 2020 Massachusetts Institute of Technology 5Las Cumbres Observatory 6California Polytechnic State University 7University of Colorado at Boulder 8University of Iowa

a rashied.amini@jpl.nasa.gov c 2019 California Institute of Technology. Government sponsorship acknowledged. 1 Executive Summary Servicing is a legal requirement for WFIRST Autonomy is the ability of a system to achieve and the Flagship mission of the 2030s [5], yet goals while operating independently of exter- past and planned demonstrations may not pro- nal control [1]. The revolutionary advantages vide sufficient future heritage to confidently meet of autonomous systems are recognized in nu- this requirement. In-space assembly (ISA) is cur- merous markets, e.g. automotive, aeronautics. rently being evaluated to construct large aper- Acknowledging the revolutionary impact of au- ture space telescopes [6]. For both servicing and tonomous systems, demand is increasing from ISA, there are questions about how nominal op- consumers and businesses alike and investments erations will be assured, the feasibility of teleop- have grown year-over-year to meet demand. In eration in deep space, and response to anomalies self-driving cars alone, $76B has been invested during robotic operation. from 2014 to 2017 [2]. In the previous Planetary The past decade has seen a revolution in Science Decadal, increased autonomy was identi- the access to space, with low cost launch ve- fied as one of eight core multi-mission technolo- hicles, commercial off-the-shelf technology, and gies required for future missions [3]. programs that have enabled numerous cubesat The impact of autonomous systems on our missions. NASA and academic institutions will ability to observe the universe can be just as be operating more small satellites and opera- revolutionary [4]. However, relevant autonomy tions centers will need to adapt. The need will work to date has been limited in scope and too be greater if future human exploration goals to disjoint to confidently deliver anticipated capa- launch dozens of cubesats per SLS launch is bilities, like in-space assembly (ISA), in a low met [7]. Operating autonomous observatories risk and repeatable manner in the 2020s or even provides one solution to this impending prob- the 2030s. This paper includes the following so lem. Notably, several ground-based observato- that the astrophysics community can realize the ries, like Las Cumbres and ALMA observato- benefits of autonomous systems: ries, have begun using autonomous operations to command large arrays of telescopes, identi- • A description of autonomous systems with fying advantages for observatories that follow relevant examples their example. Planet and presumably SpaceX’s • Enabled and enhanced observations with Starlink, private space mission operators, have autonomous systems reached a break point where traditional com- manding is inadequate to command their large • Gaps in adopting autonomous systems constellations and are operating spacecraft with automated scheduling [8]. • Suggested recommendations for adoption by Gehrels/Swift is an inspiring example of the the Astro2020 Decadal time-domain observations that autonomous sys- As we consider the observations necessary tems enable. The multi-messenger approach for to answer new science questions formed in the characterizing the physics leading to and result- 2010s, the need for autonomy is clear. Concept ing from gravity wave events will require sim- studies for the Astro2020 Decadal require opera- ilar missions to Gehrels/Swift. Gehrels/Swift tions that are more complex than ever before. relies on prescriptive state machines, statically- Increasingly complex space- and ground-based programmed conditions and routines also used observatories have more systems, components, in spacecraft fault protection, to execute au- and software. More engineering complexity in- tonomous Gamma-ray burst (GRB) follow-up variably means that there are more paths for observations. The system autonomy approach anomalies to disrupt a system’s ability to per- detailed in this paper offers several advan- form its mission. This can reduce observational tages over state machines in terms of dynamic efficiency and potentially negate the advantages decision-making and scalability. One major ad- of larger apertures and more sensitive detectors. vantage is the ability to make decisions using on-

1 missions, for instance through a cost cap credit. Adoption in the 2020s will reduce the risk of fu- ture Flagship servicing missions. 2 Understanding Autonomous Systems Observing the proceedings of the Space Astro- physics Landscape in 2020 and Beyond meeting, it is clear that a gap exists between the expec- tations of the astrophysics community and the technical readiness of autonomy technologies re- quired to meet these expectations. To under- Figure 1: Effective, reliable autonomous systems stand this gap, we need to first define autonomy must coordinate between the resources utilized by a in a relevant context. system’s lower level functions to achieve system-level goals. AppendixA offers an illustrated example of a A hierarchy of systems is represented in Fig- system autonomy framework. ure1. At the bottom of the hierarchy is the functional-level, where control and autonomy is board analysis of data to change an observation exercised in a limited domain. Functional control program. is the commanding actuators and sensors, e.g. a Dynamic decision-making also enables the command is sent and a motor turns at a com- restoration of functionality in the event of an manded rate. Functional autonomy is decision- anomaly. This type of decision-making is en- making within the boundaries of the functional abled by on-board health monitoring software, element. A simple example is a state machine which monitors and diagnoses hardware anoma- that (dis)engages a heater based on thermome- lies to support autonomous systems. This re- ter input. A more complicated example is an sults in greater observational efficiency and uni- attitude controller that takes inputs of attitude versally benefits all observatories. For observa- knowledge (e.g. star trackers). Its output is con- tories with competed time, this means more PIs trol system actuation to maintain a desired atti- can be supported. For mapping missions, like tude. Pre-programmed routines filter inputs and the Galaxy Evolution Probe, Probe of Inflation evaluate conflicting knowledge, resulting in pre- and Cosmic Origins, and Cosmic Dawn Intensity dictable behavior. Mapper Probe, greater depths can be reached More complex forms of functional auton- per unit time [9, 10, 11]. For time-domain sur- omy have already been demonstrated and are veys, this results in less gaps in data. currently being developed. For instance, au- As evidenced by private investments and de- tonomous optical navigation determines devia- velopments in ground-based observatories, the tion from desired orbit ephemeris and has been adoption of autonomous systems in space is in- used on Deep Space-1, Deep Impact/EPOXI, evitable. There are two questions to the field: other planetary missions, and soon Arcsecond “When will we start using it?” and “How will Space Telescope Enabling Research in Astro- we start using it?” Given the ambitions of physics (ASTERIA) [12, 13]. On-going work the community, the time to begin is now. In on servicing and ISA utilizes computer vision as order to use it in a repeatable, low risk, and a knowledge source to control robotic actuation cost-effective way, NASA, spacecraft vendors, [6]. On-orbit robotic servicing was first demon- and the astrophysics community need to coop- strated on DARPA’s OrbitalExpress in 2007 [14]. eratively develop a coherent technical path for- In the next few years, RESTORE-L will be used ward. To do so, our primary recommenda- to service Landsat-7 in low earth orbit using tele- tion is for NASA to incentivize the use of operation after autonomous docking [15]. autonomous systems for competed space However, functional elements utilize system

2 resources, e.g. time, power, attitude, data stor- age, etc. Spacecraft are resource limited and efficient use is critical to mission success. Dif- ferent activities may utilize resources in a mu- tually exclusive way; for instance, a space tele- scope may not be able to point its telescope at a target while simultaneously pointing its antenna toward Earth for communications. Some re- sources are zero-sum but accommodating of mul- tiple spacecraft goals; for instance, all powered equipment require power but not all subsystem power modes can be supported simultaneously. Thus, there is a state of competition between different system goals. In the current state of practice, this competition is resolved by human planning during operations. Tools are used to define system activities, like observing and trans- mitting data, based on commands that are tied Figure 2: Task networks offer numerous pathways in to certain resources. The goals of the scientists time and state-space to achieve goals requested from to observe the sky and goals of the engineers to ground operators. Implementing tasknet-based com- preserve the spacecraft are merged using these manding enables “push button-get science” missions. tools to develop time-ordered sequences of com- mands that are uplinked to the spacecraft, e.g. plished at the system-level through on-board [16]. An extension of time-ordered sequences is planning and execution. This approach contrasts conditional sequencing, where sequences use con- with traditional commanding with sequences ditional statements as a state model. This ap- through its use of task networks (tasknets, proach has the capability of storing pre-defined though goal and constraint networks are also routines that can later be executed [17]. used in the literature), described in Figure2. Autonomy poses a challenge to operational Tasks are defined as commands that are associ- planning: how can you command a system that ated with metadata defining the state conditions makes its own decisions? State machine-based required for their execution and state impacts autonomy is predictable in well-defined environ- that result from their execution. Thus, graph ments, and so resource budgets can be allotted networks of tasks can be constructed with tasks because the input domain is well characterized. as nodes and edges connecting tasks whose state Spacecraft health is further ensured by fault pro- impacts are the state requirements of another tection state machines, adding another layer of task. Moreover, tasks can be have temporal con- protection. The use of state machines enables straints to be sequence-like. In this manner an Gehrels/Swift to detect GRBs with the wide- autonomous system can be commanded like a field Burst Alert Telescope and slew to observe traditional system if desired. with its two other payloads [18]. However, ma- Sets of tasks can be defined as independent, chine learning-based decision-making and vari- uniquely prioritized system goals. Some goals able environments, exemplified by computer identify system state transitions, such as the ac- vision-guided robotic control, means that re- quisition of new science data. Other goals iden- source utilization cannot be effectively bounded tify states that need to be maintained and re- in advance and so reliable, safe operation cannot stored if lost, such as those related to spacecraft be readily assured with traditional commanding. health. The role of on-board planning and exe- Coordination of resources used by functional cution is to negotiate between the constraints of elements, prescriptive or not, can be accom- all goals so that they can be executed without

3 conflict or in violation of safe resources limits. vehicles [28]. AppendixA offers an example of The final level of hierarchy at the top of Fig- how these software are implemented in practice. ure1 supports multiple autonomous systems in Las Cumbres Observatory and Atacama Large a multi-agent architecture. Millimeter/submillimeter Array (ALMA) are ex- 3 Examples of Relevant Autonomous amples of ground-based observatories whose op- Systems erations are autonomously planned and exe- cuted. Las Cumbres Observatory is a network Spitzer Space Telescope, Dawn, Juno, and many of 18 telescopes at six sites that operate as a other planetary missions have made use of condi- single observatory, enabling persistent observa- tional sequencing with the Virtual Machine Lan- tion. Scientist request observations, which are guage (VML). Spitzer reported several advan- assigned and scheduled through a global sched- tages over traditional sequencing using VML. In uler [29, 30]. General-purpose software has been particular, it made observations contingent on developed for autonomous telescope operations telescope settling state rather than sequenced that can be adopted by future observatories op- time, which added one or two extra observations erating on these principles [31]. ALMA dynami- 1 in an 11 2 hour observing window. It also had the cally schedules and executes 30 minute “schedul- advantage of reducing spacecraft safing due to ing blocks” based on weather, science priority, on-board memory overflow [19]. However, as re- project completion, and other parameters [32]. ported in [19], the limiting factor in implement- Automated scheduling has traditionally been ing more of these autonomous behaviors was that used for operational planning. Most relevantly, there was “no fast and effective way of modeling Space Telescope Institute uses SPIKE for plan- the flight system behavior on the ground.” ning Hubble observations [33]. Planet uses au- Autonomous systems relying on on-board tomated scheduling to operate its fleet of earth planning and execution are beginning to see in- observing cubesats. In human spaceflight Time- creased use on space- and ground-based observa- liner has seen significant use on-board the Inter- tories. A prominent example is the use of AS- national Space Station (ISS) and is being con- PEN/CASPER on Earth Observer-1, which used sidered as a candidate for the Lunar Gateway on-board science planning and execution to de- [34]. While automated planning streamlines op- tect novel terrestrial scenes, like disasters, to au- erations, it still has drawbacks when the plan tonomously perform follow-up observations [20, cannot succeed due to operational conditions. 21]. Extending the work of CASPER, the Intel- ligent Payload EXperiment (IPEX) cubesat ex- 4 Astrophysics with Autonomous ecuted one year of autonomous payload opera- Systems tions using its on-board planner [22,4]. PLan 4.1 Enabled Missions and New Science Execution Interchange Language (PLEXIL) is funded to be used to for a technology demon- Autonomous systems have already enabled new stration mission of multi-agent autonomy. Later astrophysics. Both ground- and space-based in 2019, ASTERIA will demonstrate the use of transient event observatories are fundamentally the Multi-mission EXECcutive (MEXEC). Next enabled by autonomous systems. Autonomous year, Mars 2020 will use the Onboard Scheduler transient event detection and follow-up obser- to maximize science return by using excess time vation capability has been demonstrated with and power at the end of each Martian sol to plan Gehrels/SWIFT and the Zwicky Transient Fa- additional measurements [23, 24, 25]. Temporal cility [35]. As exemplified by LCO, time-domain planning and scheduling systems also include Ix- astronomy observations, e.g. supernovae, mi- TeT [26], used for robotic contorl, and EUROPA crolensing, near earth asteroids, tidal disruption [27]. Other systems have been developed based events, gravitational wave events, etc. require on similar principles since then, notably IDEA real-time, highly reactive telescope scheduling. and T-REX, used for autonomous underwater In these cases, observations cannot be planned

4 in advance and the configuration of the observa- sure). To accomplish the former goal, a robotic tions may need to evolve over time according to arm moves, changing the spacecraft’s moment of the characteristics of event. inertia. To accomplish the second goal, the at- With the projected improvements to ground- titude control system maintains attitude based based detection and localization of gravity wave on a model of the spacecraft’s moment of iner- (GW) events, there is a need for observato- tia. If robotic action is not coordinated, the at- ries that can rapidly observe potential multi- titude controller’s moment of inertia model will messenger signals. Ground-based observatories not be consistent with reality. This may lead to will require the ability to respond to external over/under actuation of reaction wheels, poten- signal, verify observability of the GW ellipse tially leading to collision risk and mission failure given current observatory conditions, and re- for both spacecraft. task to perform GW follow-up observation while Multi-agent autonomy also enables new obser- maintaining knowledge of the past observation. vations. The AEON Network is ground-based Space-based observatories will be required to facility currently under development operating do the same while also maintaining spacecraft numerous telescopes that will allow astronomers health. Gehrels/SWIFT itself was launched in to submit requests for observation in real-time. 2004 and may need replacement in the 2020s to Through multi-agent autonomy, a large network retain the community’s ability to perform GRB of ground- and space-based observatories, like detection and localization over large areas of the AEON, can coordinate their observing programs sky. If ESA’s Theseus is selected for M5, system across multiple facilities and wavelengths, serv- autonomy software and expertise can serve as a ing as a powerful tool for characterizing new dis- potential NASA contribution to that mission. coveries. Multi-agent autonomy can also be used Given the past priorities of the Astrophysics on a constellations of low cost satellites as dis- Decadal and NASA funding, it is expected tributed transient event, namely GRB, observa- that space-based time-domain observatories will tories [37, 38]. By using low cost scintillating de- be competed and are subjected to cost cap. tectors on low cost smallsat/cubesat platforms, For instance, the Gravitational-Wave Ultravio- localization can be performed through time-of- let Counterpart Imager (GUCI) has already been arrival similar to the Interplanetary Network. proposed for the SmallSat call [36]. These mis- One advantage of this approach is the timeli- sions can be architected using state machine au- ness of observation. For instance, in simulations tonomy, following the Gehrels/SWIFT. Given of flooding event observations by an earth ob- the bounded nature of time-domain observations serving constellation, a multi-agent architecture and the additional advantages afforded by on- measured flood area to 96% accuracy over time board planning/execution, we note that these as opposed to a centrally planned architecture missions can alternatively use on-board plan- observing with 70% accuracy, owing to the time- ning/execution as a relatively low risk means of liness of observation [39]. Additionally, multi- demonstrating and increasing the community’s agent coordination of more than two assets may confidence in the technology. be required, or would greatly facilitate, interfer- As discussed in [6], system-level autonomy is ometry missions such as LISA. required for ISA and servicing in order to co- A unique class of missions that would bene- ordinate robotic autonomy with the rest of the fit from ISA and multi-agent autonomy are ra- spacecraft. One example of how critical system- dio and possibly NIR/optical/UV interferometry level autonomy is to ISA and servicing is the co- missions that may require ISA of large apertures ordination of a mass model as robotic operation and coordination. is performed. At a high-level, a servicer space- Autonomous systems also complement the in- craft has the goals of performing robotic oper- creased access to space afforded by small satel- ation and assuring attitude control in the pres- lites and low-cost launch vehicles. As the total ence of disturbance (gravity gradient, solar pres- number of missions increases, let alone missions

5 that may utilize more than one spacecraft such exozodiacal light is not so bright that it re- as GUCI, the ground stations and operations duces the effective raw contrast at the exo- facilities become a bottleneck for commanding planet’s location. Even if exozodical light is and monitoring spacecraft. At some point, large previously characterized in mid-IR wavelengths numbers of traditional spacecraft cannot be effi- [42], these observations may not predict the ex- ciently commanded through traditional means. ozodiacal light at HabEx/LUVOIR near UV to Autonomous systems reduce the human effort near IR wavelengths. Additionally, not all sys- required to command as the burden can be off- tems will have constrained inclination that im- loaded to an on-board planner. pacts the apparent brightness of the exozodia- Current plans for human exploration offer new cal dust. Currently, HabEx and LUVOIR will platforms for astrophysics missions, creating new schedule their observations in advance and use a opportunities for the development of observato- pre-determined observing program based on lit- ries. Most imminently, lunar exploration may tle or no knowledge of the actual level of exozodi create dozens of opportunities for new measure- optical brightness around individual targets. ments. With cubesats piggybacking launches In an autonomous system, coronagraphic and opportunities to use the Lunar Gateway as imaging can be analyzed on-board the spacecraft a platform for payloads, managing multiple mis- to evaluate the contribution of exozodiacal light sions and scheduling observations that may have and the determine the value of continuing obser- conflicting pointing and thermal requirements on vation. In this case, excessive exozodical light Lunar Gateway becomes increasingly difficult to can be detected on-board within a fraction of coordinate across multiple teams [40]. Again, the planned observation time. On-board data an operational bottleneck results that can be re- processing software can then alert the on-board solved through automated planning. Addition- planner to truncate the observation so the ob- ally, returning to the moon creates new opportu- servatory can perform the next scheduled obser- nities for lunar surface-based observatories. This vation. Data from the truncated observation is offers unique opportunities for some radio bands, later downlinked for future analysis. In this ex- cosmic ray, MeV γ-gay, X-ray, and UV measure- ample, more targets are observed more quickly, ments that cannot be made from Earth’s sur- resulting in more observing time for other targets face. A Probe mission concept, FARSIDE, is a of interest and greater exo-Earth yield during the ∼10 MHz radio observatory on the farside of the primary mission [43]. Moon. As it requires a rover for deployment, Recommendation: NASA should use autonomous mobility and robotic assembly ca- ROSES as a means of funding software de- pability is critical to mission feasibility. [41] velopment for on-board data processing. 4.2 Efficient Observing Programs The advantage of on-board data processing The traditional paradigm of commanding re- in union with a system planner is not limited duces the overall efficiency of targeted ob- to space-based observatories. Subsystems that serving programs as observation length is pre- evaluate weather and seeing conditions can aid to determined in advance. Later, data is down- autonomously reschedule planned observations linked and analyzed on the ground. However, that may not be possible when scheduled, im- the efficiency of observing programs can be im- proving their net efficiency. proved by analyzing data on-board to inform sys- Recommendation: NASA and NSF tem decision-making. should incentivize the development of future One example is exoplanet direct imaging, ex- ground-based observatories with automated emplified by HabEx and LUVOIR, that requires scheduling/execution, following the example a level of 10−10 raw contrast to perform di- of ALMA and LCO. rect imaging of exo-Earths. This raw contrast can only be effectively achieved in cases where As discussed above, on-board data processing

6 and multi-agent autonomy can be used in a co- ordinated network of ground- and earth-based to perform GW follow-up observatories. In such an architecture, localization that is currently per- formed post-hoc as data is released can be per- formed on-board within the constellation, result- ing in localization while the source is still emit- ting brightly. 4.3 Adaptive Fault Protection Enables More Observations Figure 3: Histogram of safing events binned on the Traditional space systems have fault protection number of days between suspension and restoration schemes that enter safe modes, requiring human of nominal operation. With on-board planning and diagnosis and commanding to restore nominal execution and on-board health diagnosis, about 50% operation. As a result, 4% of nominal space- of anomalies resulting safing may be averted. Result flight operations are blocked by spacecraft saf- based on analysis of the [44] safing dataset. ings [44]. Notably, [44] presents a lower bound prove ground- and balloon-based observatories. on blocked operational time, as other anomalies Ballooning in particular suffers from a high fail- can occur that restrict nominal operations and ure rate, owing from ad hoc integration of mul- do not cause safing. tiple payloads on-site and schedule constraints Figure3 indicates that on-board plan- forcing limited testing. Recently, NASA JPL ning/execution with on-board health diagnosis evaluated technologies for a self-reliant rover may mitigate the impact of about 50% of saf- during which on-board health diagnosis was ings. This adds an additional week of nom- found to be effective in the build, integration, inal operations per year. There are two rea- and testing environments in discovering and di- sons. Rather than relying on state machines for agnosing hardware issues previously undetected executing fault protection, health maintenance [49, 50, 51]. Health diagnosis software can be tasknets can restore the minimum functionality used for ballooning systems that are used re- required to perform science operations while not peatedly, such as mirror motor control, pressure endangering spacecraft health [45, 46]. Second, vessels, and power generation, to detect hard- this architecture permits integration of on-board ware issues. This can reduce complexity and health diagnosis to monitor the health of hard- stress during the balloon integration phase and ware and local models for attitude knowledge improve success rate of balloon missions. and control, a major cause of safing events, to in- form these health maintenance tasknets [47, 48]. Recommendation: Integrate the use of While an additional week of data per space health diagnosis software for elements that observatory may seem marginal, if applied to are repeatedly used on ballooning platforms. NASA’s fleet of space-based observatories it would result in 11 additional weeks of science 5 Addressing the System Autonomy per year for the community. The benefit is most Gap useful to observatories with PI-directed observa- tions, like Hubble and Spitzer, where additional For astrophysics, autonomous systems can en- PIs can be supported. For mapping missions, able and enhance missions that deliver revolu- like GEP, additional mapping depth per unit tionary data sets, reduce the cost of missions, time is achieved. For time-domain surveys, cov- and reduce the burden on scientists in developing erage is more complete in time. and maintaining observing programs. A future Even without on-board planning/execution, where “press button − get science” missions is health diagnosis models and software can im- on the horizon, but work remains that requires

7 the community’s awareness and support. restrictive to software technologies that can be The primary gap is cultural. Autonomous effectively validated outside of the operational systems imply a different paradigm of de- environment, e.g. on-board science data process- sign and operations compared to traditionally- ing software. commanded systems. Compared to autonomy Recommendation: NASA should evaluate in the private sector, a small proportion of our the applicability of the Technology Readiness space science and spaceflight communities have Level as a means of evaluating the maturity relevant expertise to review the opportunities of autonomy and on-board data processing and risks associated with autonomous systems. software. This is compounded by traditional engineering preference for heritage designs and expectations Other gaps are technical. Autonomy frame- of predictability. On point, how can scientists, works, described in AppendixA, define rules for engineers, and proposal reviewers be confident in how system-level planning and execution inter- a mission concept that operates itself? Is it pos- face with traditional components and systems sible to design and deploy autonomous systems and functional autonomy. Community accep- that are partially autonomous to placate the con- tance of these frameworks can reduce adoption cerns of the community? These questions need risk and promote repeatability by permitting tra- to be formally addressed if NASA is to meet its ditional design and operations approaches. By legal requirement to perform servicing require- defining a convention for how autonomous mis- ment for future large, space-based observatories, sions should be designed and built, frameworks let alone to reap the benefits of autonomous sys- would also improve reviewability of autonomous tems for observation. missions and portability of testing methodology. Limited institutional capacity to adopt au- Remaining work includes improving the verifi- tonomous systems is exemplified by the exam- ability of tasknets, which is critical to reaping ples of autonomous systems above: most of these the benefits of integrated fault protection. Re- missions were or will be designed and built by latedly, telemetry that permits reconstruction of NASA. Given the high cost and risk associated on-board decision-making requires further study with changing the process by which spacecraft and definition. Ground systems and tools for are designed, built, and tested, spacecraft ven- commanding of autonomous spacecraft require dors have till now relied on conditional sequenc- further maturation. ing and not autonomous planning/execution. Finally, some observations will benefit from Thus, government-industry cooperation is re- on-board data analysis. For these observations, quired to make use of autonomous systems re- new software will be required to perform this liably and repeatably for all NASA missions. function, which will be the responsibility of sci- ence community. While not the subject of this Recommendation: NASA should incen- white paper, processing-intensive data process- tivize the use of autonomous systems for ing may require high performance computing. competed space missions. Specifically, High performance computing does not necessar- small sat missions, missions of opportunity, ily enable autonomous systems, but is enhanc- SMEX, MIDEX, and Probe missions can in- ing by permitting intensive processing of science clude a credit for using the technology. We data and on-board scheduling over larger search note that transient event observatories offer spaces. a low risk path to maturing this critical tech- nology. Recommendation: NASA should fund technology demonstrations of high- Another aspect of the cultural gap is NASA’s performance space-based computing for definition of technology readiness and its refer- on-board data processing. ence to “operational environment” that is overly

8 A Example of a System Autonomy under development at NASA SSC; and, a vehi- Framework cle management system for autonomous space- craft habitat operations is under development at In order to create autonomous systems repeat- NASA ARC and JSFC [13, 52, 53]. Finally, the ably and reliably, a framework must be de- European Robotic Goal-Oriented Autonomous fined. Similar to a legal constitution, a sys- Controller (ERGO) has been under developed tem autonomy framework defines responsibili- by an EU-funded consortium of industry and ties and capabilities of system components and academia [54]. Below, we use FRESCO as an how they interface with one another in govern- example to illustrate how systems autonomy is ing system behavior. For instance, the man- implemented. ner in which on-board science data processing is Tasknet Tasknets are data structures that en- interfaced to inform decisions at a system-level capsulate the potential envelop of spacecraft be- should be identical across all missions regardless havior. They are graph networks where nodes of which decisions it informs. This enables re- are tasks and the edges are the state and tempo- liable mutli-mission adoption of on-board plan- ral dependencies between tasks. Tasknets can be ning/execution. Under a unified framework, en- defined as goals for the spacecraft to achieve, to gineers, scientists, and managers can work to- transition states (e.g. an imaging survey goal re- ward the same set of requirements that assure sults in a set of images being taken) and to main- mission success. Reviewers can also use the same tain states (e.g. a pointing knowledge mainte- framework to verify compliance. Without a uni- nance goal restores pointing knowledge through fied framework, development, mission assurance, optical navigation if a knowledge uncertainty and review can become intractable given the threshold is violated). Tasknets have been de- complexity in designing autonomous systems. scribed in literature since the 1970s [55, 56]. There are several requirement that define an Planner and Executive Planners create and effective autonomy framework. It should make maintain schedules of tasknets based on their guarantees about acceptable behavior, enable prioritization and projected timelines of future confident operator oversight and insight, readily states. Scheduling tasks is performed by a search accommodate new information, and not require function, whose search space can be constrained extensive tailoring or ad hoc modification to sup- based on how tasks are defined. This permits port multiple missions. Pragmatically, such a traditional, sequence-like behavior or highly au- framework must afford a practical path toward tonomous behavior within the same framework. adoption. To do so, system autonomy must in- At a certain time before scheduled execution, the tegrate with existing components. Human work- planner passes tasks to the executive for execu- flows involved across mission phases should de- tion. Executives are responsible for intelligent viate minimally from existing practice. Also, the execution and monitoring the impact of executed framework must support varying degrees of au- tasks. Under nominal operation, they receive re- tonomy – permitting sequence-like commanding ceipt of successful task execution and proceed to to highly autonomous operation. Without this dispatch the next scheduled tasks. If a task fails, practical path, NASA and industry partners will they can exercise contingency behaviors speci- have to invest in brand new software and pro- fied by the task, which can include replanning cesses and accept significant risk in implement- requests to the planner. ing a major leap toward systems autonomy at Currently, MEXEC and PLEXIL are main- once. tained by NASA JPL and ARC, respectively [24, There are several examples of such a frame- 57]. A wider survey of command execution sys- work. The Framework for Robust Execution and tems is presented in [58]. Scheduling of Commands On-Board (FRESCO) State Database A state database serves as is under development at NASA JPL; the NASA a “single source of truth” for the system, main- Platform for Autonomous Systems (NPAS) is taining component status and abstracted system

9 Figure 4: An example of an autonomous system framework, the Framework for Robust Execution and Scheduling of Commands On-Board (FRESCO), defines capabilities and interfaces that will resulting in repeatable and predictable implementations of systems autonomy for complex systems, such as spacecraft. This figure offers a simplified description of FRESCO components and interfaces. states used in decision-making. terfaces will vary. For instance, if a hardware System-Level Estimator Estimators that controller includes local fault protection, an in- inform decision-making use system telemetry as terface for a signal to interrupt system-level exe- input to models of system behavior. System cution over that controller’s domain is required. health monitoring software is one such estima- Traditional hardware controllers and on-board tor. System health monitoring serves two pur- data processing can also be used to inform the poses. First, it can be used to identify poten- scheduling of tasknets. The on-board processing tially faulty components to alert operators to po- for exozodical light in exoplanet coronagagraphy tential future risks. In rover testing, it was able serves as an example. to successfully identify undiagnosed hardware B Acknowledgements problems[51]. Second, it permits the creation Thank you to Ellen Van Wyk, NASA JPL, for of tasknets that operate only if healthy compo- illustrations included in this white paper and nent states are reported, reducing the risks of SpaceX for the cover photograph. autonomous operation. MONSID, developed by Okean Solutions, uses linked models of hardware behavior to monitor the health status of compo- nents [48, 59, 51]. Function-Level Software and Compo- nents Function-level software can perform a multitude of functions, ranging from hardware control to data processing and functional auton- omy. Depending on its function, its internal in-

10 References [11] T-C Chang for the CDIM Team. “CDIM [1] NASA Autonomous Systems – Systems Cosmic Dawn Intensity Mapper”. In: The Capability Leadership Team. Autonomous Space Astrophysics Landscape for 2020 Systems Taxonomy. 2019. and Beyond. Apr. 2019. [2] CF Kerry and J Karsten. Gauging invest- [12] Deep Space-1. Navigation: Primary Mis- ment in self-driving cars. url: %5Curl % sion. JPL: DESCANSO Design and Per- 7Bhttps : / / www . brookings . edu / formance Summary Series, Apr. 2004. research / gauging - investment - in - [13] Lorraine Fesq and Rashied Amini. “A self-driving-cars/%7D. New Paradigm for Autonomous Space- [3] Space Studies Board, National Research craft: From Research to Deployment”. In: Council, et al. Vision and voyages for plan- Low Cost Planetary Missions (LCPM-13). etary science in the decade 2013-2022. Na- 2019. tional Academies Press, 2012. [14] T Weismuller and M Leinz. “GN&C tech- [4] Steve Chien and Kiri L Wagstaff. “Robotic nology demonstrated by the orbital express space exploration agents”. In: Science autonomous rendezvous and capture sen- Robotics 2.7 (2017), eaan4831. sor system”. In: 29th annual AAS guid- ance and control conference. American As- [5] Science Committee on Commerce and tronautical Society. 2006. Transportation. National Aeronautics and Space Administration Authorization Act of [15] R Ticker. Restore-L Mission Information. 2010: report of the Committee on Com- url: %5Curl%7Bhttps://www.nasa.gov/ merce, Science, and Transportation on S. sites / default / files / atoms / files / 3729. U.S. G.P.O., 2010. restore - l - info _ nnh15heomd001 _ r7 . pdf%7D. [6] Rudranarayan Mukherjee et al. “When is it worth assembling observatories in [16] P Maldague et al. “APGEN: A multi- space?” In: (July 2019). mission semi-automated planning tool”. In: First international NASA workshop on [7] Kimberly F Robinson and Andrew Schorr. planning and scheduling. 1998, pp. 363– “NASA’s Space Launch System: Excep- 365. tional Opportunities for Secondary Pay- loads to Deep Space”. In: 2018 AIAA [17] Christopher Grasso and Patricia Lock. SPACE and Astronautics Forum and Ex- “VML sequencing: Growing capabilities position. 2018, p. 5138. over multiple missions”. In: SpaceOps 2008 Conference. 2008, p. 3295. [8] V Vitaldev. “Automated Scheduling and Operation of a Heterogeneous Fleet of [18] D Palmer. GRB Follow Up Observation Satellites”. In: IWPSS 2019. 2019. Autonomy. In correspondence. June 2019. [9] Jason Glenn et al. “The Galaxy Evolution [19] David S Mittman and Robert Hawkins. Probe: a concept for a mid and far-infrared “Scheduling Spitzer: the SIRPASS story”. space observatory”. In: Space Telescopes In: (2013). and Instrumentation 2018: Optical, In- [20] S. Chien et al. “ASPEN - Automating frared, and Millimeter Wave. Vol. 10698. Space Mission Operations using Auto- International Society for Optics and Pho- mated Planning and Scheduling”. In: In- tonics. 2018, p. 106980L. ternational Conference on Space Opera- [10] S Hanany. “PICO Probe of Inflation and tions (SpaceOps 2000). Toulouse, France, Cosmic Origins”. In: The Space Astro- June 2000. physics Landscape for 2020 and Beyond. Apr. 2019.

11 [21] Daniel Tran et al. “The autonomous sci- [29] Nikolaus Volgenau and Todd Boroson. encecraft experiment onboard the EO- “Two years of LCOGT operations: the 1 spacecraft”. In: Proceedings of the challenges of a global observatory”. In: fourth international joint conference on Observatory Operations: Strategies, Pro- Autonomous agents and multiagent sys- cesses, and Systems VI. Vol. 9910. Inter- tems. ACM. 2005, pp. 163–164. national Society for Optics and Photonics. [22] Steve Chien et al. “Onboard autonomy 2016, p. 99101C. on the Intelligent Payload EXperiment [30] T Boroson et al. “Science operations for (IPEX) CubeSat mission: A pathfinder for LCOGT: a global telescope network”. In: the proposed HyspIRI mission intelligent Observatory Operations: Strategies, Pro- payload module”. In: Proc 12th Interna- cesses, and Systems V. Vol. 9149. Inter- tional Symposium in Artificial Intelligence, national Society for Optics and Photonics. Robotics and Automation in Space, Mon- 2014, 91491E. treal, Canada. 2014. [31] RA Street et al. “General-purpose soft- [23] Lorraine Fesq et al. “Extended Mis- ware for managing astronomical observing sion Technology Demonstrations Using the programs in the LSST era”. In: Software ASTERIA Spacecraft”. In: 2019 IEEE and Cyberinfrastructure for Astronomy V. Aerospace Conference. IEEE. 2019, pp. 1– Vol. 10707. International Society for Op- 11. tics and Photonics. 2018, p. 1070711. [24] Vandi Verma et al. “Autonomous Sci- [32] A Bridger. “The ALMA Observing Tool: ence Restart for the Planned Europa Mis- Proposal Preparation & Submission”. In: sion with Lightweight Planning and Ex- ALMA Early Science Workshop, IRAM. ecution”. In: International Workshop on ALMA. 2010. Planning and Scheduling for Space (IW- [33] Mark D Johnston. “Spike: Ai schedul- PSS 2017), Pittsburgh, PA. 2017. ing for nasa’s hubble space telescope”. In: [25] Gregg Rabideau and Ed Benowitz. “Proto- Sixth Conference on Artificial Intelligence typing an onboard scheduler for the mars for Applications. IEEE. 1990, pp. 184–190. 2020 rover”. In: Proceeding of Interna- [34] PK Barnes, AT Haddock, and CA Cruzen. tional Workshop on Planning and Schedul- “Autonomous Science Operations Tech- ing for Space, Pittsburgh, PA. 2017. nologies for Deep Space Gateway”. In: [26] Philippe Labone and Malik Ghallab. Deep Space Gateway Concept Science “Planning with sharable resource con- Workshop. Vol. 2063. 2018. straints”. In: Proceedings of the 14th inter- [35] Roger M Smith et al. “The Zwicky national joint conference on Artificial in- transient facility observing system”. In: telligence. Vol. 2. Citeseer. 1995, pp. 1643– Ground-based and Airborne Instrumenta- 1649. tion for Astronomy V. Vol. 9147. Interna- [27] Javier Barreiro et al. “EUROPA: A tional Society for Optics and Photonics. platform for AI planning, scheduling, 2014, p. 914779. constraint programming, and optimiza- [36] Stephen B Cenko. “The Gravitational- tion”. In: 4th International Competition on wave Ultraviolet Counterpart Imager Knowledge Engineering for Planning and (GUCI) Network”. In: American As- Scheduling (ICKEPS) (2012). tronomical Society Meeting Abstracts. [28] Conor McGann et al. “T-rex: A model- Vol. 234. 2019. based architecture for auv control”. In: 3rd Workshop on Planning and Plan Execution for Real-World Systems. Vol. 2007. 2007.

12 [37] Judith Racusin. “Current and Future [46] Jeremy D Frank and Gordon B Aaseng. Large Missions as Science Motivation for “Transitioning Autonomous Systems GRB Detecting SmallSats”. In: Towards a Technology Research to a Flight Software Network of GRB Detecting Nanosatellites. Environment”. In: AIAA SPACE 2016. 2018. 2016, p. 5530. [38] Rashied Amini. “Utilizing Fractionated [47] Ksenia Kolcio, Louis Breger, and Paul Space Mission Design and Small Satellites Zetocha. “Model-based fault management for a Next Generation Gamma Ray Burst for spacecraft autonomy”. In: 2014 IEEE Observatory”. In: 22nd SmallSat Confer- Aerospace Conference. IEEE. 2014, pp. 1– ence. 2008. 14. [39] Sreeja Nag et al. “Autonomous Schedul- [48] Ksenia Kolcio and Lorraine Fesq. “Model- ing of Agile Spacecraft Constellations with based off-nominal state isolation and de- Delay Tolerant Networking for Reactive tection system for autonomous fault man- Imaging”. In: 11th International Workshop agement”. In: 2016 IEEE Aerospace Con- on Planning and Scheduling for Space. ference. IEEE. 2016, pp. 1–13. 2019. [49] Daniel Gaines et al. “Productivity chal- [40] William Gerstenmaier and Jason Cru- lenges for mars rover operations”. In: The san. Cislunar and Gateway Overview. url: 26th International Conference on Auto- https://www.nasa.gov/sites/default/ mated Planning and Scheduling. 2016. files/atoms/files/cislunar-update- [50] Daniel Gaines et al. “Productivity chal- gerstenmaier-crusan-v5a.pdf. lenges for mars rover operations”. In: The [41] Jack Burns, Gregg Hallinan, et al. “CDIM 26th International Conference on Auto- Cosmic Dawn Intensity Mapper”. In: The mated Planning and Scheduling. 2016. Space Astrophysics Landscape for 2020 [51] Ksenia Kolcio, Ryan Mackey, and Lorraine and Beyond. Apr. 2019. Fesq. “Model-Based Approach to Rover [42] B Mennesson et al. “Constraining the Health Assessment-Mars Yard Discover- exozodiacal luminosity function of main- ies”. In: 2019 IEEE Aerospace Conference. sequence stars: complete results from the IEEE. 2019, pp. 1–12. Keck Nuller mid-infrared surveys”. In: The [52] Fernando Figueroa, Mark Walker, and Astrophysical Journal 797.2 (2014), p. 119. Lauren W Underwood. “NASA Platform [43] Christopher C Stark et al. “Lower limits for Autonomous Systems (NPAS)”. In: on aperture size for an exoearth detect- AIAA Scitech 2019 Forum. 2019, p. 1963. ing coronagraphic mission”. In: The Astro- [53] Richard Levinson et al. “Development and physical Journal 808.2 (2015), p. 149. Testing of a Vehicle Management System [44] Travis Imken et al. “Modeling space- for Autonomous Spacecraft Habitat Oper- craft safe mode events”. In: 2018 IEEE ations”. In: 2018 AIAA SPACE and As- Aerospace Conference. IEEE. 2018, pp. 1– tronautics Forum and Exposition. 2018, 13. p. 5148. [45] Gordon B Aaseng, Adam Sweet, and John [54] J Oc´on et al. “The ERGO framework Ossenfort. “Performance Analysis of an and its use in planetary/orbital scenarios”. Autonomous Fault Management System”. In: Proc. 69th International Astronautical In: 2018 AIAA SPACE and Astronautics Congress (IAC), IAF, Bremen, Germany. Forum and Exposition. 2018, p. 5149. 2018.

13 [55] Richard E. Fikes and Nils J. Nilsson. “Strips: A new approach to the applica- tion of theorem proving to problem solv- ing”. In: Artificial Intelligence 2.3 (1971), pp. 189–208. url: %5Curl%7Bhttp://www. sciencedirect.com/science/article/ pii/0004370271900105%7D. [56] SA Vere. “Deviser-An AI planner for spacecraft operations”. In: (1985). [57] Vandi Verma et al. “Universal-executive and PLEXIL: engine and language for ro- bust spacecraft control and operations”. In: Space 2006. 2006, p. 7449. [58] Vandi Verma et al. “Survey of command execution systems for NASA spacecraft and robots”. In: (2005). [59] Ksenia Kolcio, Lorraine Fesq, and Ryan Mackey. “Model-based approach to rover health assessment for increased productiv- ity”. In: 2017 IEEE Aerospace Conference. IEEE. 2017, pp. 1–13.

14