Sensor Webs and Multiple Agents
Sep. 12th, 2010 02:29 pmThe project I'm currently working on involves the programming of multiple satellites to work in coordination using multi-agent techniques. Steve Chien, from JPL, seems to be one of the key people when it comes to getting artificial intelligence technology onto satellite systems. Have gave about 10 talks at iSAIRAS. I'm not sure if that is typical of the field or a result of various belt-tightening measures (he wasn't the first author on many of the papers he was presenting and actually commented at one point on the difference in style between his various sets of slides). Something similar happened at this year's AAMAS where Medhi Dastani gave about six talks because his institution had refused to fund PhD students to travel to the conference.
Two of Chien's talks involved cooperation between a mixture of satellites and on-ground sensor systems, or robots. One was a very speculative piece of work ultimately aimed at seismic and atmospheric events (e.g. dust devils) on Mars. The practical work involved a small rover robot and a couple of mounted cameras in a constructed "Mars Yard" which could be coordinated to make observations.

Picture of the JPL Mars Yard
The other talk discussed an existing Volcano monitoring sensor web in which involves ground sensors at Mount St. Helen's (and other networks) which can request observations from the EO1 Satellite. This is already deployed and indeed observations were automatically triggered during the Icelandic Volcano eruptions this year. The Volcano web already uses Multi-agent technology so seems very relevant to our work.
EDIT: One thing that we (at least the Liverpool end of the project) hadn't really clocked to, but which became very obvious listening to Chien's talks was that all these satellites have a complex schedule of observations they have to make. These present quite hard planning problems since the observations are constrained not only by the time the satellite is over the right bit of the world but also by data storage, uplink and downlink times and bandwidths and instrument heating. We've been talking about cases where a group of satellites need to move into some configuration in order to make an observation or, alternatively, where one satellite malfunctions (or one of its instruments malfunctions) and they have to change formation to compensate. Clearly we now need to at least think about how such reconfigurations would effect the large scale planning process, as well as the immediate observation at hand.
Two of Chien's talks involved cooperation between a mixture of satellites and on-ground sensor systems, or robots. One was a very speculative piece of work ultimately aimed at seismic and atmospheric events (e.g. dust devils) on Mars. The practical work involved a small rover robot and a couple of mounted cameras in a constructed "Mars Yard" which could be coordinated to make observations.

Picture of the JPL Mars Yard
The other talk discussed an existing Volcano monitoring sensor web in which involves ground sensors at Mount St. Helen's (and other networks) which can request observations from the EO1 Satellite. This is already deployed and indeed observations were automatically triggered during the Icelandic Volcano eruptions this year. The Volcano web already uses Multi-agent technology so seems very relevant to our work.
EDIT: One thing that we (at least the Liverpool end of the project) hadn't really clocked to, but which became very obvious listening to Chien's talks was that all these satellites have a complex schedule of observations they have to make. These present quite hard planning problems since the observations are constrained not only by the time the satellite is over the right bit of the world but also by data storage, uplink and downlink times and bandwidths and instrument heating. We've been talking about cases where a group of satellites need to move into some configuration in order to make an observation or, alternatively, where one satellite malfunctions (or one of its instruments malfunctions) and they have to change formation to compensate. Clearly we now need to at least think about how such reconfigurations would effect the large scale planning process, as well as the immediate observation at hand.