How does a robot know when something goes wrong? Our research answers this question by leveraging expectations - predictions about the immediate future – and using the mismatch between the expectations and the external world to monitor the robot’s progress. We use the cognitive architecture ACT-R (Adaptive Control of Thought - Rational) to learn the associations between the current state of the robot and the world, the action to be performed in the world, and the future state of the world. These associations are used to generate expectations that are then matched by the architecture with the next state of the world. A significant mismatch between these expectations and the actual state of the world indicate a problem possibly resulting from unexpected consequences of the robot’s actions, unforeseen changes in the environment or unanticipated actions of other agents. When a problem is detected, the recovery model can suggest a number of recovery options. If the situation is unknown, that is, the mismatch between expectations and the world is novel, the robot can use a recovery solution from a set of heuristic options. When a recovery option is successfully applied, the robot learns to associate that recovery option with the mismatch. When the same problem is encountered later, the robot can apply the learned recovery solution rather than using the heuristics or randomly exploring the space of recovery solutions. We present results from execution monitoring and recovery performed during an assessment conducted at the Combined Arms Collective Training Facility (CACTF) at Fort Indiantown Gap.

Date Published: 17 May 2013
PDF: 8 pages
Proc. SPIE 8741, Unmanned Systems Technology XV, 87410P (17 May 2013); doi: 10.1117/12.2016167

Published in SPIE Proceedings Vol. 8741:
Unmanned Systems Technology XV
Robert E. Karlsen; Douglas W. Gage; Charles M. Shoemaker; Grant R. Gerhart, Editor(s)