The next generation of autonomous systems—from robots and drones to industrial platforms and intelligent vehicles—depends on Physical AI: systems that continuously learn, predict, and adapt to the physical world. These capabilities increasingly rely on digital twins, computational models that reconstruct and predict system behavior in real time. However, existing approaches often depend on cloud computing or computationally intensive algorithms that are too slow and energy-demanding for deployment on edge devices.
Researchers at Arizona State University have developed MERINDA, a novel framework that enables real-time digital twins and physics-guided AI models to run efficiently on resource-constrained hardware. MERINDA combines a highly parallel neural architecture with FPGA acceleration to replace traditional iterative solvers used in model recovery and neural differential equation methods.
By dramatically reducing runtime, memory usage, and energy consumption while maintaining state-of-the-art accuracy, MERINDA enables autonomous systems to continuously reconstruct and predict their dynamics directly on embedded platforms. The technology supports fast, adaptive, and trustworthy operation in mission-critical environments where cloud connectivity, latency, and power consumption are major constraints.