kwc.org Photos Spare Cycles MythBusters

Talk: AI and the New Exploration Vision

Dan Clancy, NASA

I enjoyed this talk -- it was a survey of NASA's current AI-based missions, including current and future Mars missions (Sojourner/Spirit/Opportunity). metamanda would have liked at least one point the talk made, which was that NASA is working on a personal rover robot to create/inspire kids, and in particular, girls. They have found in their exhibits that robots are more engaging to girls than boys, who enjoy the embodied interaction, so they see in it an opportunity to bridge a gender gap as well as inspire a future generation in NASA's vision. It was interesting how the Personal Exploration Rover pictures really did look like baby versions of the Spirit/Opportunity rovers, i.e. there was a certain amount of anthropomorphism to the vehicle, and it appeared child-like that could help engender a care-taker relationship between a kid and the robot.

Read on for notes.

NASA needs intelligent, multi-domain systems. Need them to accomplish vision of improving life here, extending life there, and to find life beyond.

NASA's current plan is stepping-stone-based, rather than destination-based. Don't have the same budget they used to have.

Humans "on site" enable discoveries. Vision is human + robots together in exploration. Apollo missions leveraged human intelligence.

Role of lunar exploration: dev core capabilities, operational experience, science experiments, commercial potential

Architectural and algorithmic innovation: cross-cutting theme in AI.

Remote Agent

  • Ames/JPL collaboration to put an AI system into space
  • significance is what was not accomplished.
    • Unified architecture model, but no underlying canonical interpretation.
    • Model-based programming as software engineering, a lot of hacking in what the system should do.
    • All or nothing approach = nothing.
    • Ground controllers need to be tightly coupled with system

Mars

Sojourner

  • total distance: 100M
  • max distance from Lander: 12M
  • 45-75% time waiting (e.g. b/c something went wrong)
  • Lander stopped working because of battery failed due to thermal cycle, could have gotten more out of Pathfinder itself

MER

MER was second robotic mission to go do mars.

It takes MER a day to do what a geologist can do in 45 seconds (Steve Squyers)

Very little of Mars explored, limited by where we are capable of going. * Spirit traveling toward Columbia hills. A lot of volcanic material on surface so mostly just traveling. * Opportunity discovered evidence of water. Now in Endurance crater with a broken wheel, driving robot backwards to drag wheel, saving wheel for portions of mission when they might need it.

Yankee Go Home!

MER tech

  • Descent image motion estimation system. Initial assumption about variance in wind was wrong, but as it was descending would take three images that it could use to estimate horizontal velocity and could fire jets to eliminate velocity. In analyzing photos, had to use limited computation resources to pick features that would be present in all three images for horizontal velocity estimate. If this system had not been present, one of the rovers would not have successfully landed.
  • Onboard capability for obstacle avoidance, but still some drift, so still takes roughly 3-4 days to go 10M and place instrument (successive approximation).
  • Autonav: goes about a 1M, takes new images with hazcam, analyzes, etc...
  • Uses stereo-based visual odometry
  • Has gone 90M, even though 50M of actual visual distance in images
  • Planner for scientific activity. Two months before mission the operators were dubious about using it (question of new technology adoption). Really liked declarative constraints, even outside of the actual planner. 20-50% increase in science return from planning system. People who do sequence generation say they cannot do job without system. Changed the work practice. Temporal flexibility. Mixed-initiative interaction. Allows for time ranges of activities, keeps track of impact of relations and restrictions on other activities.
  • 6-12 times when plan broke on MER, but that was because plans were very conservative.

Mars Science Laboratory 2009

  • nuclear power (Planned radioisotope power source)
  • 600-1000 day mission. complete changes way of controlling mission because people cannot live on Mars time for that length of time the way they did for MER.
  • expected distance in 10's of kms. Much greater accessible terrain.
  • Controlled propulsive landing. Considering lowering rover from crane (1m/s) as spacecraft ascends.

Desired features: * consistent semantics throughout the system (MER had ~5 different views of the world -- temporal flexibility wasn't encoded throughout system) * balance between on-board computation and pre-compilation on the ground * validation of plans, * integrating it into the system development cycle, * fault protection (robust dynamic replanning, hardware failures, resource utilization). * would like for on-board decision making for safe replanning so plans can be more aggressive. * uncertainty is a key challenge

Exploration missions

  • humans and robots together. Still have not figured out what this means yet. Will drive more human-centered systems (social/cognitive/sensory).
  • Habitat, launch vehicle, orbital vehicle need to work together.
  • Apollo example: instead of astronauts talking to each other, one talks to Earth control (everything controlled from Earth), then Earth control talks back to the other astronaut. Can't do that with Mars, too far.

Crew-centered operations a grand challenge for AI. * wearing mobile agents. will need to understand context so that system will be aware of what astronaut is currently doing. have to integrate data from multiple sources (GPS, digital camera, video camera, science instruments, speech input) * robot vehicles * habitat * system will have to integrate data and move to appropriate location (e.g. transfer photos back to habitat) * RIALIST dialogue system: innovation is the entire assembled system. Interpret speech in context of dialogue. Figure out whom is being spoken to over an open mic.

Personal Satellite Assistant

  • Free-flying robot for zero g
  • 8 camera vision, 6 degrees of control
  • integrate mobile and immobile software agents
  • mobile applications: inventory tracking, sensory calibration

Robonaut

  • upper-humanoid body
  • dexterous
  • stereo vision , human tracking, object tracking/recognition
  • MIT working on hybrid system for action interpretation and motion control (combination of symbolic and logic-based reasoning). build declarative model of what you expect human to be doing, then look at human and try to diagnose what human is currently doing (continuous models). Translate high-level tasks into low-level motion control.

Multi-agent control architecture * explicit coordination between layers and market-based task allocation. * three-tiered architecture with communication at lower levels. use bidding system to determine who is going to do what with what. behavioral level, executive level, planning level.

Sub-vocal control

Trying to create tight coupling between human and machine, extension of human senses.

Working on sub-vocal control. Place sensors on body. Can disambiguate between 30-40 words.

Neural net classification.

Working on detecting consonants and vowels.

Personal Exploration Rover

Space is cool, robots are cool, robots in space are way cool

Want to inspire children, create public interest

"Personal computer" robotics

Robotics are engaging to children

70% of girls who stand around exhibit go and do it (boys only at 60%). Robotics engage imagination and can cross gender problem. Embodied in physical world, stimulates creativity.

Conclusions

Human/robotic exploration a grand challenge for AI and intelligent systems * AI is pervasive around us, but it is often only in pieces. Need to create integrated systems (architectural innovation instead of algorithmic)

Exploration offers an opportunity for truly integrated intelligent systems * architectural and algorithmic innovation * human-centered systems * transition to software engineering practice * continued integration of symbolic and probabilistic approaches

Post a comment


tags.

related entries.

what is this?

This page contains a single entry from kwc blog posted on July 27, 2004 10:09 AM.

The previous post was My niche.

The next post is Talk: Real Robots for the Real World.

Current entries can be found on the main page.