VLA Models for Real-Time Robotics: A Comparative Study of Transformer and SSM-Based Architectures

Vision-Language-Action (VLA) models have emerged as a promising paradigm for generalist robotic control, enabling robots to translate visual observations and natural language instructions directly into executable actions. Nevertheless, their deployment in real-world robotic systems remains constrained by strict real-time requirements, particularly regarding inference latency and memory footprint. To address these limitations, recent research has explored State Space Models (SSMs), such as Mamba and Mamba-2, as an alternative to the Transformer backbones that dominate current VLA architectures. By compressing sequence history into a fixed-size latent state, SSMs enable linear-time sequence modeling with a constant memory footprint, making them attractive candidates for real-time robotic applications. This thesis investigates whether these theoretical advantages translate into practice, analyzing the trade-off between computational efficiency and task performance in Transformer-based and SSM-based VLA architectures on consumer-grade hardware. The work begins with a theoretical and architectural analysis of the two backbone families, followed by an empirical evaluation under a unified protocol on the LIBERO simulation suite. The evaluation extends beyond success rates to key real-time control metrics, such as inference latency, achievable control frequency, and memory consumption across different quantization settings. Taken together, these results offer an empirical indication of the extent to which SSMs can be a viable backbone for real-time and resource-constrained generalist robot control.