Towards Metacognitive Robot Decision Making for Tool Selection

The capability to self-asses our performance before doing a task is essential for the decision making process, e.g., when selecting the most suitable tool for a given task. While this form of awareness has been identified in humans as metacognitive performance (thinking about the performance), robots still lack this cognitive ability. This awareness has a potential to enhance their embodied decision power, robustness and safety. Here, we take a step in this direction by proposing a novel synthetic model that unites active inference with some ideas from metacognition. We (mathematically) identify three main components that contribute to the agent’s self-evaluation when making a decision: i) its performance for task completion, ii) its control effort towards task completion, and, very importantly and novel, iii) its self-confidence about the decision. We further show that these quantities are seamlessly balanced inside the free energy objective. As a proof of concept, we framed our theoretical account within the tool selection problem as a use case. Results show that the agent is able to select the best tool—modelled as spring-mass-damper systems—given three types of control tasks: attain a goal position, velocity and acceleration. Interestingly, the proposed tool selection criteria prioritises the performance during a hard task, and self-confidence during an easy task. Furthermore, we discuss how our mathematical framework can be generalized for tool/model optimization and invention.