Error Recovery in Robot Systems

Citation

Srinivas, Sankaran (1977) Error Recovery in Robot Systems. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/P7QY-KN31. https://resolver.caltech.edu/CaltechTHESIS:09102018-150127337

Abstract

This dissertation addresses itself to the problem faced by a robot in recovering from failures during execution of a task. Failures occur partly because sensory information is inaccurate, partly because effectors do not always perform as expected, and partly because the domain in which the robot operates cannot be characterized exactly. Robot systems with automated planners have traditionally dealt with the problem of error recovery by merely replanning to achieve the desired goal, without attempting to characterize the failure in any way whatsoever.

The central idea in this thesis is that planning recovery from failures has its own special techniques, distinct from those used in conventional planning systems. Two viewpoints, looking at the past for an explanation of the failure, and looking at the current situation to attempt a characterization of the failure state, provide powerful heuristics for error recovery. This thesis suggests that these heuristics can be formalized as failure reason analysis and multiple outcome analysis, and that knowledge relevant for such analysis can be provided through a failure reason model and a multiple outcome model associated with each action.

The failure reason model about why actions provides a means for representing fail, like bumping into an object to be grasped because of servoing errors or because of inaccurate information about the location of the object. The model also provides knowledge required for distinguishing between the different reasons for failure. Finally, it includes recommendations of corrective actions to be taken if failure is attributed to a specific reason. This model in used in failure reason analysis in building a failure tree representing possible explanations of the failure. The explanations represented in the tree are then used in planning recovery.

The multiple outcome model provides a way of representing the possible outcomes of an action, like bumping onto the object or bumping onto the ground in the immediate vicinity of the object, ignoring the fact that these outcomes could be the result of several different reasons. Knowledge required to distinguish between different outcomes is provided as part of the model. In cases where the immediately available information is inadequate to identify the outcome of an action, the multiple outcome model provides a basis for executing actions to serve as information gathering steps. The novel feature here is that information gathering is directed by specific expectations about the state of the world.

A computer implementation of a program called MEND has provided a medium for exploring the above idea. MEND has been designed to automate recovery from failures in simple manipulation tasks to be performed by the JPL robot, but the techniques used in MEND have greater generality. A first implementation of MEND established the basis of this investigation. A second version, which has been designed to correct some limitations of the first version, has not yet been fully implemented and integrated with the JPL robot system.

The techniques of planning recovery from failures through failure reason analysis and multiple outcome analysis are contributions to the subject of robotics. More importantly, however, the problem of error recovery is recognized to be a member of a larger class of problems involving knowledge representation and common sense reasoning, both of which are core topics in the study of artificial intelligence. The solution presented in this thesis makes some new contributions to these core topics.

Thesis Files