Monthly Technical Report

March 1999


We have divided this report into two portions. The first portion discusses the site visit on March 17 by John Blitch and Douglas Gage and the second portion describes our technical progress in the month of March.

SITE VISIT

John Blitch and Douglas Gage visited our lab on March 17. During this visit, we introduced the members of our group to them, described the experimental set-up we have assembled during the past three months in our lab, and demonstrated some of our software running on the robots, as well as other algorithms running in simulation. After their visit, they requested us to prepare videos that demonstrate
  1. the range data gathered by the SICK laser as the robot moves,
  2. the different trade-offs in the next-best view algorithm for map-building,
  3. how our target-finding algorithm in 2D adapts to different properties of the sensor (specifically, omnidirectional sensors versus sensors with cone vision), and
  4. our target-finding algorithm for an aerial observer.
We are currently preparing these videos. They will be ready shortly.

One of the concerns raised by John Blitch during the visit was how our algorithms will generalize to three-dimensional environments. There are two avenues for such generalizations.

  1. As we explained in our report on environment, perception, and mobility models, we represent the environment by a set of 2D plans (each plan corresponds to the floor of a building). The plans are connected to each other by portals (portals model staircases, elevators, and escalators). It is straightforward to extend our next-best algorithm to construct such 2-1/2D models, provided that the next-best view algorithm is supplied with software to recognize staircases, elevators, and escalators. Further, it is also simple to extend our target-finding and target-tracking algorithms to work with such representations.
  2. We can also extend our algorithms to deal with 3D obstacles on each floor, where we model each obstacle as a prism. In the next stage of the map-building task, we intend to construct 3D models of the environment. In order to do so, we plan to mount the range sensor vertically on the robot so that each scan of the sensor corresponds to a vertical slice of the environment.
We can also adapt our target-finding and target-tracking techniques to work in an environment with such obstacles. In fact, our target-finding algorithm for an aerial observer operates in precisely such an environment! We plan to use the expertise developed in the aerial-observer project to extend our target-finding and target-tracking techniques to environments with prismatic obstacles.

TECHNICAL RESULTS

We are also making steady progress towards our milestones for the Third and future Quarterly IPR meetings. In particular, we are refining the implementation of our next-best view technique for building 2D maps to handle larger uncertainties in robot positions. We have implemented robust algorithms for processing and fusing the data returned by the laser range sensor. We are improving the performance of these algorithms by executing them on actual range data collected by the laser range sensor.

We have also completed the implementation of a basic target-tracking algorithm on our SuperScout robots equipped with pan-tilt cameras. This algorithm enables the tracker to maintain a constant distance from a moving target. We are currently implementing a more soophisticated target-tracking planner that uses the model of the environment constructed in the map-building phase in order to track the target more robustly. Our algorithm computes how quickly the target can move out of the visibility region of the tracker and moves the tracker in such a way as to maximize the escape time of the target.