INFORMATION INVARIANTS FOR DISTRIBUTED MANIPULATION

In Donald (1995), we described a manipulation task for cooperating mob ile robots that can push large, heavy objects. There, we asked whether explicit local and global communication between the agents can be rem oved from a family of pushing protocols. In this article, we answer in the affirmative. We do so by using the general methods of Donald (199 5), analyzing information invariants. We discuss several measures for the information complexity of the task: (I) How much internal state sh ould the robot retain? (2) How many cooperating agents are required, a nd how much communication between them is necessary? (3) How can the r obot change (side effect) the environment to record state or sensory i nformation for performing a task? (4) How much information is provided by sensors? and (5) How much computation is required by the robot? To answer these questions, we develop a notion of information invariants . We develop a technique whereby one sensor can be constructed from ot hers by adding, deleting and reallocating I) through 5), among collabo rating autonomous agents. We add a resource to measures I) through 5) and ask: 6) How much information is provided by the task mechanics? By answering this question, we hope to develop information invariants th at explicitly tradeoff resource 6) with resources I) through 5). The p rotocols we describe here have been implemented in several different f orms, and we report on experiments to measure and analyze information invariants using a pair of cooperating mobile robots for manipulation experiments in our laboratory.