Factor graph represents the factorization of a probability distribution function and serves as an effective abstraction in various autonomous machine computing tasks. Control is one of the core applications in autonomous machine computing stacks. Among all control algorithms, Linear Quadratic Regulator (LQR) offers one of the best trade-offs between efficiency and accuracy. However, due to the inherent iterative process and extensive computation, it is a challenging task for the autonomous systems with real-time limits and energy constrained. In this paper, we present FGLQR, an accelerator of LQR control for autonomous machines using the abstraction of a factor graph. By transforming the dynamic equation constraints into least squares constraints, the factor graph solving process is more hardware friendly and accelerated with almost no loss in accuracy. With a domain specific parallel solving pattern, FGLQR achieves 10.2x speed up and 32.9x energy reduction compared to the software implementation on an advanced Intel CPU.

翻译：因子图表示概率分布函数的因式分解，并作为自主机器计算任务中的有效抽象。控制是自主机器计算堆栈中的核心应用之一。在所有控制算法中，线性二次型调节器（LQR）在效率与精度之间提供了最佳权衡。然而，由于固有的迭代过程及大量计算需求，对于具有实时限制和能量约束的自主系统而言，这是一个具有挑战性的任务。本文提出FGLQR——一种利用因子图抽象实现自主机器LQR控制的加速器。通过将动态方程约束转化为最小二乘约束，因子图求解过程对硬件更友好，且加速后几乎不损失精度。凭借领域特定的并行求解模式，与在先进Intel CPU上的软件实现相比，FGLQR实现了10.2倍的加速和32.9倍的能耗降低。