HW/SW Acceleration of VLA Models for Robotics on Mesh-of-Tiles Architectures

Vision-Language-Action (VLA) models are emerging as a unifying paradigm for robotic autonomy, mapping camera streams and language instructions directly into motor commands. On embodied platforms their transformer backbones must run within closed-loop latencies of tens of milliseconds and a few watts. Mesh-of-tiles accelerators such as MAGIA provide the compute density required, but exploiting them demands HW/SW co-design.This Incarico di Ricerca, aligned with the ARCHYTAS project, focuses on the HW/SW acceleration of VLA and autonomous-navigation models on mesh-of-tiles architectures, with MAGIA as reference platform. The incaricato di ricerca will focus on: 1) mapping of VLA workloads onto the mesh (operator partitioning, tiling, scheduling, quantization); 2) hardware extensions and dataflow support for the dominant operators; 3) software stack (compiler, runtime, kernel libraries, host-to-fabric dispatch) and end-to-end validation of a closed-loop navigation task on FPGA.