Beyond Action Repeat: Characterizing the Precision-Horizon Tradeoff and Internalizing Temporal Resolution in Skill Discover

Embodied agents typically make control decisions at fixed temporal resolutions, selecting new actions at regular intervals. However, different tasks may favor different temporal resolutions. Temporal abstraction is one way to address this problem, allowing the agent to select an action or skill that persists across multiple environment steps. This creates a precision-horizon tradeoff that has been noted empirically yet has been studied only over narrow temporal ranges and in low-dimensional settings. Existing unsupervised skill learning methods typically treat skill execution duration as an external hyperparameter rather than explicitly training skills to fit a fixed temporal resolution. In this work, we characterize the precision-horizon tradeoff on robotics tasks spanning from two dimensions to twenty-eight dimensions, where excessively low temporal resolutions lose precise control while excessively high temporal resolutions inflate the horizon. We also found that the optimal action repeat curve is task-dependent. We further explore the precision-horizon tradeoff by modifying Contrastive Successor Features (CSF)---an existing unsupervised skill learning algorithm---to learn skills with an intrinsic length. Our results show that the precision-horizon tradeoff persists for learned skills and not just primitive actions, the skill length $L$ needs to be appropriately matched with an action repeat $k$, and that our modification improves CSF in our evaluations.