Abstract
<title>Abstract</title> <p>Autonomous robots must determine not only which actions to perform, but also what issues should be considered before, during, and after task execution. However, existing systems commonly rely on fixed reasoning procedures or allow an LLM to organize the entire reasoning process independently for each task, resulting in omitted or redundant thinking activities. This paper proposes a continual and open-ended thinking activity scheduling framework that learns what a reasoning component should think about rather than how the reasoning itself should be performed. A background program continuously collects thinking-related events from previous LLM calls, task execution histories, environmental observations, acquired knowledge, failures, and successful exploratory activities. These records are organized as sequential learning data to schedule thinking purposes such as cause analysis, handling strategy generation, future impact analysis, and action determination. The framework supports cross-task learning, allowing thinking scheduling knowledge acquired from different physical tasks to be combined for previously unseen compound tasks. Its thinking activity space is also extensible: new activities can be introduced through semantic descriptions and prompt templates, and their scheduling conditions can subsequently be learned. A thinking intervention mechanism is further provided to examine and revise provisional schedules when necessary. Results show that the learned scheduler maintains a 96.6% task success rate while reducing LLM-related calls by 61.9% compared with a fixed sequence, improves unseen-task F1 from 0.701 to 0.805 through cross-task learning, and achieves a new-activity F1 of 0.846 while reducing operational violations from 60.5% to 15.8%.</p>