AI in Space Lessons From NASAs Deep Space

AI in Space – Lessons From NASA’s Deep Space 1 Mission Ron Keesing Dept. Of Maths & Computing Science University of the South Pacific

Overview • • • NASA’s Deep Space 1 Mission A brief introduction to AI Why use AI in space? The Remote Agent Future Directions

NASA’s Deep Space 1

NASA’s Deep Space 1 Technology • First mission of NASA’s New Millenium Program. • Primary mission goal was to demonstrate new technologies.

NASA’s Deep Space 1 Innovation First demonstration of • Ion Propulsion System Engine • AI system for spacecraft command control (Remote Agent) • 10 other technologies

NASA’s Deep Space 1 Ion Propulsion System Engine

NASA’s Deep Space 1 Exploration • Launched in October, 1998 • Successful flybys of – Asteroid Braille (26 Kms) • Closest asteroid approach to date – Comet Borrelly • Best data ever collected from comet

NASA’s Deep Space 1 Exploration – Launch

NASA’s Deep Space 1 Exploration – Artist’s conception

NASA’s Deep Space 1 Exploration - Encounter with Comet Borrelly

A Brief Introduction to AI • What is artificial intelligence? – Creating machines that behave “intelligently”. • How would you know an “intelligent” machine if you saw one? – The Turing Test

A Brief Introduction to AI • Early predictions (1950 s) were very optimistic. • Computers would soon: – Win World Chess Championship. – Understand spoken language. – Be as “smart” as people.

A Brief Introduction to AI • HAL-9000 from “ 2001 – A Space Odyssey”

A Brief Introduction to AI • HAL-9000 from “ 2001 – A Space Odyssey”

A Brief Introduction to AI • HAL-9000 from “ 2001 – A Space Odyssey”

A Brief Introduction to AI • Creating “intelligent” computers has proven far more difficult than imagined. – 50 years to beat world chess champion. – Not close to understanding language. – No one even thinks about a computer passing the Turing Test anymore…

A Brief Introduction to AI • Human beings are much smarter than we gave ourselves credit for. – Solving differential equations is easy. Catching a ball is hard. – Storing lots of information is easy. Understanding it is hard.

A Brief Introduction to AI • Even though AI is hard, we’ve developed a lot of ways to make machines “smarter”. – Planning – Automated Reasoning – Agents

A Brief Introduction to AI • Planning – Decomposing high-level goals into individual tasks that can achieve these goals efficiently • Robots that can manipulate objects. • Systems for optimizing manufacturing processes.

A Brief Introduction to AI • Automated Reasoning – Making inferences from limited information based on knowledge of the domain. • Medical diagnosis • Theorem proving

A Brief Introduction to AI • Agents – Systems that can perform complex tasks autonomously. • Web search agents that can navigate the web looking for specific pieces of information.

Why Use AI In Space? • Conventional model of spacecraft control – Ground control sends a plan. – Spacecraft executes that plan. – If anything goes wrong, spacecraft enters “safe mode” and calls ground control.

Why Use AI In Space? • Problems with “ground control” approach: – Expensive – May miss opportunities. – Sometimes it’s more dangerous to wait for help. – Some missions can’t be run from ground.

The Remote Agent • An autonomous system for spacecraft command control • 3 components – Planner – Inference system (MIR) – “Smart Executive”

The Remote Agent • Planner – Takes mission goals from ground control and creates a plan to satisfy them • Plans for limited resources (power). • Plans satisfy numerous constraints (orientation, communication, state of devices).

The Remote Agent • Planner Op. Nav IPS Thrust ACS Turns MICAS Power MICAS Imaging Nav OD Planning

The Remote Agent • Inference system (MIR) – Gets information from sensors. – Deduces state of spacecraft using modelbased reasoning. – Suggests ways to reconfigure if devices fail.

The Remote Agent • Inference system (MIR)

The Remote Agent • “Smart Executive” – Sends commands to spacecraft to execute the plan. – Executes the plan flexibly, including trying multiple methods if necessary. – Recognizes when a plan has failed and triggers replanning.

The Remote Agent • “Smart Executive” (to_achieve (IPS_THRUSTING ips level) ((ips_is_in_standby_state_p ips) (sequence (achieve (power_on? 'ega—a)) (command_with_confirmation (send—ips—set—thrust—level)) (command_with_confirmation (send—acs—change—control—mode : acs—tvc—mode)))) ((ips_in_thrusting_state_p ips) (command_with_confirmation (send—ips—change—thrust—level))) (t (fail : ips_achieve_thrusting)))

The Remote Agent Experiment • The Remote Agent took control of DS-1 for 3 days. – Validated all objectives • Demonstrated ability to formulate plans to satisfy mission goals. • Demonstrated ability to diagnose faults and reconfigure to perform tasks. • Demonstrated plan failure, recovery, and replanning.

The Remote Agent Experiment • Lessons from using AI on DS-1. – An autonomous system can successfully control a complex spacecraft. • Opens the door to new types of missions and a new relationship between ground and spacecraft. – The RA architecture is a powerful approach to building robust autonomous systems

Future Directions For AI In Space • • Autonomous rovers Manned mission to Mars Formation flying Many-agent approaches