DPMA and the AERCam Testbed

The AERCam simulation was a testbed for Neodesic's Dynamic Predictive Memory Architecture (DPMA) for integrating task execution (using RAPs) and natural language (using the Conceptual Memory Parser, CMP).

The DPMA testbed is implemented using Macintosh Common Lisp, and uses Apple's QuickDraw 3D, Speech Synthesis and Speech Recognition

Important Issues in DPMA

Resolving natural language descriptions to particular actions and objects
Conversational plans
Shared control between human operator and autonomous agent
Self narration of activities
Language generation
Describing active tasks

Representative Language

Here's a transcript that's representative of natural language interaction with DPMA.

Pictures

Screenshot #1 (236K)	Screenshot #2 (150K)
Screenshot #3 (184K)	Screenshot #4 (202K)
AERCam free flyer	AERCam point of view (70K)
Orbiter occupancy grid for navigation in configuration space, 36 inch resolution. Orbiter occupancy is approximate. (86 K)	Satellite occupancy grid for navigation in configuration space, 36 inch resolution. Satellite occupancy is approximated with a sphere. (136K)
Example of 3D path planning: A path from a point observing the nose cap to a point observing the main engine cluster. 3D path planning can be interesting; I cheated by doing best-first search in the occupancy grid and doing some post-processing to reduce the number of waypoints. To learn more about the real thing, see "Path Planning and Control for AERCam, a Free-flying Inspection Robot in Space" by H. Choset and D. Kortenkamp (available from the CMU Robotics Institute).

John Wiseman / wiseman@neodesic.com
May 7, 1999