COM-PACT: COMponent-Aware Pruning for Accelerated Control Tasks in Latent Space Models

The rapid expansion of resource-constrained mo- bile platforms, such as mobile robots, wearable systems, and Internet-of-Things devices, has heightened the need for com- putationally efficient neural network controllers (NNCs) that can operate within strict hardware limitations. Although deep neural networks (DNNs) achieve high performance in control applications, their considerable computational complexity and memory demands hinder practical deployment on edge de- vices. This study presents a comprehensive model compres- sion methodology that employs component-aware structured pruning to determine the optimal pruning magnitude for each group, thereby balancing compression and stability for NNC deployment. The proposed approach is rigorously evaluated on Temporal Difference Model Predictive Control (TD-MPC), a leading model-based reinforcement learning algorithm, with systematic integration of mathematical stability guarantees, specifically Lyapunov criteria. The principal contribution is a principled framework for identifying the theoretical limits of model compression while preserving controller stability. Experimental results indicate that the methodology effectively reduces model complexity while maintaining essential control performance and stability. Additionally, the approach defines a quantitative boundary for safe compression ratios, enabling practitioners to systematically determine the maximum per- missible model reduction before compromising critical stability properties, thus supporting the reliable deployment of com- pressed NNCs in resource-limited environments.