UC San Diego

Industrial robots have remained the fastest-growing market, within the overall robotics industry sales, in recent times. These robots have shown that when tasked with repetitive, structured tasks under controlled constraints they can contribute to society by ensuring consistent quality and significantly bringing down amortized costs. However, the requirements of the sector itself are moving towards a future where a high mix of customization with low-volume production of each type is of value. This "high-mix, low-volume" setting brings a number of realistic, but yet unattainable, robotics challenges which are summarized by the overarching question of how to transfer known plans for a given task to a similar environment but different in a number of ways?

We formulate this as a question of "plan repair" which revises the original plan based on new or contrary observations made by the robot during execution. We assume a nominal plan and base demonstrations for grounding it and solely focus on the problem of how to: autonomously monitor a system for failures, localize the cause of said failure, and recover by repairing the original assembly plan. The recurring theme in this endeavor has been tackling failures of different kinds and at different levels of planning which decrease the overall robustness of the system. We observed that by contextualizing failures using the knowledge of how the task progresses in nominal conditions we were able to reason about the pathology of it and implemented different recovery procedures in our execution framework to deal with them accordingly. This thesis sets out to validate the claim that: Use of meta-reasoning for failure detection, fault localization, and plan repair in assembly robotics requires grounding the meta-knowledge in action and perception. Grounded meta-reasoning for these tasks is more flexible than classical planning and less costly than situated learning.

In support of this claim we make the following contributions:1. A multi-resolution planning and monitoring system for assembly planning and execution, assuming complete knowledge 2. Methods for learning task models with incomplete procedural knowledge 3. Methods for repairing task plans with incomplete object or pose-grounding knowledge 4. A schema for a generic meta-reasoning architecture motivated by our key research insights