A Framework for Robot Self-Assessment of Expected Task Performance

Toward Competent Robot Apprentices: Enabling Proactive Troubleshooting in Collaborative Robots

“…These costs can be learned from real-time robot experience, which is a realization we will make use of. This is seen in [38] and [6], in which the robot learns success likelihoods through robot experience. In [38], this knowledge is used for a later planning problem, where the agent attempts to maximize success and efficiency.…”

“…In [38], this knowledge is used for a later planning problem, where the agent attempts to maximize success and efficiency. In [6], this is used for humanrobot dialogues about potential action success, failures, and counterfactuals. Our system will provide both of these features.…”

“…While plans are often selected by their length, we instead perform plan selection based upon a probability of success assessment: when an action or set of actions has been selected, it can be performed by various task-specific or robot-specific functions. As [6] introduces, and we have expanded, action outcomes are monitored and tracked, and so the likelihood of an autonomously generated plan can be computed. Similarly, as we have introduced and will outline, unexpected events imply the presence of a not-yet-measured state, allowing a context to be changed if necessary.…”

“…The robot tracks information about all action outcomes using the technique described in [6]. In this approach, it is assumed that the architecture has the ability to observe the success or failure of an action.…”

“…Recent work in robot self-assessments based on past task performance (e.g., [6]) is highly relevant in this regard as past experience can be used to make predictions about future behavior outcomes. But just predicting failures does not address how to overcome them or how to work with a human partner to address ongoing faults.…”