Researchers argue world models deserve study for adapting remote robot teams in the field

Jamie Santos and Kagan Tumer argue in a position paper posted on arXiv that world models should be studied as a way for autonomous robot teams in remote places to adapt after they have been deployed.

Changes in the environment, the hardware or the make-up of a team can make policies that once worked less effective, they write. A new mission goal can be met by picking another solution from an existing set of trade-offs, but new deployment conditions can reorder that set, and trying candidates on the real robots puts the hardware at risk.

World models could predict how each policy would perform under the latest estimate of conditions, so a team could re-evaluate its options without running them, as long as the predictions match what the deployed robots actually do, the authors write. They set out research priorities, including checking predictions against real results and sharing information between team members.