To study Earth and explore the vast reaches of space, NASA utilizes the recent advances of Artificial Intelligence (AI). The Jet Propulsion Laboratory AI Group performs fundamental research in fields of Artificial Intelligence Planning and Scheduling, with applications to science analysis, deep space network operations, spacecraft commanding and space transportation systems.
Today, we are going to elaborate some of the major projects that JPL is currently working on. Most of them are related to planning technology, spacecraft autonomy and rover autonomy.
Table of Contents
The AI Group at JPL has been working on a system known as ASPEN (stands for Automated Scheduling and Planning Environment). It’s a modular, reconfigurable framework that can support numerous types of planning and scheduling applications.
The framework includes a variety of components that implements the elements, mostly found in complex planning or scheduling systems, such as a temporal reasoning system, a resource management system, an expressive modeling language and a graphical interface. Specifically, it is used in mission design planning, spacecraft operations, surface rover planning, coordinated multiple rover planning and multiple rover planning.
As a ground based system, ASPEN utilizes an internal spacecraft model and different high level objectives to provide specific commands to be executed by the spacecraft. As a flight based system, it continuously receives updates on rover state and updates the plan to reflect surrounding changes. As an antenna scheduling system, it is used to autonomously control a DSN station.
The AI technology is based on heuristic search, iterative repair and temporal reasoning. The framework has a generic architecture that makes it easy to select from different propagation algorithms and search engines to make the planning process effective. Moreover, users can interact with the schedule and re-plan quickly and efficiently.
It is currently available for external licensing but not export. In future, ASPEN will be used to integrate repair planning with execution.
Multi-Rover Integrated Science Understanding System (MISUS) develops technologies to control rovers for planetary exploration. The MISUS architecture developed by NASA consists of three main components –
Data Analysis: A distributed machine learning system, performing unsupervised clustering for modeling rock type distribution observed by rovers. It can direct rover’s sensing to continually improve the content of the planetary scene.
Planning: A distributed planning system, generating operation plans to achieve input rover science goals. There is a central planner that splits up science goals between rovers, and a distributed planner set associated with each operation upon an individual rover.
Environment Simulator: A multiple rover simulator, modeling several geological surrounding and rover science operations. It deals with science data of all surroundings, keeps track of operations, and reflects observations by rover science equipment.
The overall system operates in a closed loop manner where data analysis system can be viewed as scientists driving exploration process. First, the data are transmitted to rover clustering algorithms, which integrates all collected data into an updated global model and broadcasts the new model back to the distributed clusterers.
The clustering output is used by a prioritization algorithm to produce new set of observation goals, which will further improve the model’s accuracy. The goals are then transmitted to a central planner that assigns individual rovers to goals in a way that will most efficiently process the requests.
Each rover planner then generates some specific actions which will achieve as many of its assigned goals as possible. The sequences of action are then sent to the simulator where they’re implemented and any collected information is sent back to the rover clusterers. The whole cycle continues until enough information is collected to generate distinct clusters for any observed rock types.
4. Distributed Spacecraft
The project uses latest technologies to control constellations of spacecraft with mission objectives rather than command sequences for every individual spacecraft.
This research advances the modeling and simulation capabilities to enable high-precision, real-time simulation of spacecraft formations and clusters through distributed technologies.
NASA is developing a new simulation architecture to utilize the distributed nature of the formation and split the simulation among multiple processor in a cluster. HYDRA (Hierarchical Distributed Reconfigurable Architecture), for instance, is developed to seamlessly deploy simulated modules and technologies across mixed and multi platform environments.
HYDRA automates the communication process between simulation modules. It has been successfully infused into FAST (Formation Algorithms and Simulation Testbed) as a part of Terrestrial Planet Finder program.
The overall objective is to build robust and fast global optimization algorithms that can solve formation flying guidance, estimation, control and decision making problems. This includes fast distributed estimators for formation flying, distributed resource allocation among spacecraft, robust formation keeping control, formation fuel optimal reconfiguration path planning, and mode commanding.
CASPER (stands for Continuous Activity Scheduling Planning Execution and Replanning) utilizes iterative repair to support continuous alteration or modification of spacecraft.
The conventional batch oriented planning models of planning have several shortcomings. Building a plan from scratch demands an extensive amount of computation, and onboard computational resources are typically limited.
The goal is to make planner more effective and responsive to unexpected changes. The planner that can decrease the dependency on predictive models, like inevitable modeling errors.
To achieve this, JPL utilizes a continuous planning technique what is known as CASPER. The planner has a current goal set, state, and a model of the expected outcome. An incremental update to the current state can be applied any time. This update could be anything, from simple time progressing tweaks to unexpected events.
The planner further maintains a consistent plan with the latest data available. However, most of the time, things do not go as per expectations. That’s where planner comes in action – it stands ready to continuously alter the plan according to the scenario.
Multi-Rover Execution Architecture
Current iterative repair planning approach enables incremental alterations to the initial state as well as objective and then step-by-step resolve conflicts. After each iteration, its effect will be propagated to conflicts discovered and the plan updated (for example plan repair algorithms invoked).
This technology is used in Planetary Rover Operations, New Millennium Earth Orbiting 1, Citizen Explorer, highly reusable space transpiration, Distributed Rovers, Modified Antarctic Mapping Mission, and more.
2. Volcano Sensorweb
The project uses a network of sensors linked by internet and software to an autonomous satellite observation response capability. It is developed with a modular, flexible architecture to facilitate expansion in sensors, customization of trigger scenarios and responses.
Till date, it has been used to implement a global surveillance project in order to examine volcanos. Moreover, NASA runs sensorweb tests to study cryosphere events, flooding and atmospheric phenomena.
Sensorweb Detection and Response Architecture
Several operational satellites make their data available for free, such as data from the MODIS (Moderate Resolution Imaging Spectrometer) are available in almost real time through Direct Broadcast. These data provide global and regional coverage with impressive sensing capabilities.
However, these equipments do not provide data in high resolution suitable for many science applications. In fact, most of them are high-demand assets and highly constrained.
In Volcano Sensorweb, high coverage and low resolution sensors are used to trigger observations by high resolution devices. Also, there are numerous other rationales for network of sensors into a sensorweb. For instance, automated response could make observation possible via complex devices like imaging radar. Or, they could be used to increase the observation frequency in order to enhance the temporal resolution.
For now, it’s being used to monitor 50 most active volcanos on Earth. In addition, NASA also runs experiments to monitor wildfires, flooding and cryospheric event.
The spacecraft used in the previous NASA missions (before 2000) had no ability to make autonomous decisions by themselves based on the data they collect in space. However, ASE (Autonomous Sciencecraft Experiment (ASE) that is being operated onboard on Earth Observing-1 mission since 2003, uses continuous planning, onboard pattern recognition and machine learning to increase efficiency.
The ASE software demonstrates the ability to use onboard decision-making to identify, examine and respond to events and downlink only the data that contains highest value.
This AI technology includes numerous useful modules such as
- Onboard science algorithm to analyze image data to detect trigger conditions
- Robust execution management software to enable event-driven processing and low-level autonomy.
- CASPER software to re-plan crucial activities like downlink.
ASE opens a wide range of new opportunities in Earth science, space physics and planetary science. The technology decreases the downtime lost to anomalies, reduces equipment setup time by using autonomy software, and dramatically increases the science per fixed downlink.
Initially, ASE contains science targets to monitor high level objectives. CASPER is used to generate plan for periodically monitoring targets (using Hyperion instrument). The onboard science algorithms examines the images and the images are downlinked based on its detection. If there is no suitable event, the science software commands the planner to acquire the next highest priority target.
Then, the SCL software implements the plans generated by CASPER in conjunctions with different autonomy elements and this cycle is repeated on subsequent observations.