kwc.org Photos Spare Cycles MythBusters

Talk: Real Robots for the Real World

Sebastian Thrun, Stanford

This talk was mostly over my head in terms of the math, but the work is interesting.

1992 was first AAAI visit. First robot competition. Had to navigate hexagon from pole to pole in random/predefined orders. Won 3rd place.

Don't have to go as far to look for unexplored places for robots. * mining and outdated maps. have to worry about collapsing -- mine subsidence large problem around the world, can cause small earthquakes. 9 miner in Somerset trapped when they accidentally breached into other mine that was filled with water (didn't know exactly where other mine was because of outdated maps). huge number of mines in PA area. total number of mine unknown. danger of subsidence, ground-water contamination, mine fires (can burn for decades, multiple mine fires per year)

Mine mapping systems

  • robots don't need oxygen
  • first reused robots from other projects, eventually built new robot for mine mapping. didn't have much money so had class project with students (funny video of student at keyboard while others building robot).

Mine-mapping robot about as big as a person laid flat. had to protect against sparks. combustible gas detectors. have to worry about railroad tracks, mud and water. slow moving. had to explosion-proof, but wasn't water-proof.

Amounts to a 3-D navigation task. laser-range finders (move, stop, panning laser), darkness detectors.

NO: communication (wireless does not work), human access, sparks, information on hazards (unreliable memories), breadcrumbs, second robot

Mapping/SLAM

simultaneous localization and mapping (SLAM) * two camps that don't cross-over * topological camp: relative coordinates, coarse, cognitively plausible. build map with nodes and arcs. hasn't produced any systems that have really worked yet. * metric camp: absolute coordinates, detailed, 'engineeringly' plausible. build detailed probability map. topo camp says that this isn't how people work, won't scale. * can convert between two using matrix inversion/-log probability

Mr. Metric: * measure location with Gaussian cloud. uncertainty will increase as it moves through mine. not independent uncertainties -- can reduce uncertainty of previous positions. Have to maintain correlations. Extended Kalman Filter. * have real-world results with robot in Sydney Harbor * state: map + robot * posterior: p(map, robot) * embedded in euclidean space * update time O(n^2) * proactive mapping

Ms. Topological: * only cares about relative coordinates. collects on local information, so it scales. * build information matrix with inverse Kalman filter * In topo space, if you have sparse information matrix, can update in constant time (requires approximation -- matrix approximately sparse). think very locally. * can think of connection between local information as rubberbands and can distort map as a whole by applying forces * not embedded, all relative * constant update time O(1) * state: map + robot * posterior: -log p(map, robot) * but: no inference * lazy mapping

Can convert loss-free between topological and metric spaces

Cycles are a problem with matrices. mines have many, many cycles

Ms. Data Association * data association problems rare * data association = soft constraint * real-time recursive algorithm * can undo associations * search problem, NP-hard, possibly exponential, maintain fringe of tree so can backtrack to other hypotheses

Translate 3-D scans into cost map for movement. Could find shortest path through cost function, but costs can be rotation dependent (e.g. railroad tracks)

Problem: after mine abandoned for a year roof often collapses. Need different types of robots: e.g. reconfigurable boring robot.

DARPA autonomous vehicle grand challenge * purely technological challenge. * out of 150 mile course, winning team went 7 miles * AI challenge: need to reason about terrain

Post a comment


tags.

related entries.

what is this?

This page contains a single entry from kwc blog posted on July 27, 2004 2:18 PM.

The previous post was Talk: AI and the New Exploration Vision.

The next post is Talk: AI Characters and Directors for Interactive Computer Games.

Current entries can be found on the main page.