Biologically-Inspired Hardware for Land/Aerial Robots

Broadly, this work focuses on the design of biologically inspired (e.g., artificial neural network) hardware and can be roughly broken down into three specific phases: hardware for robotic navigation, hardware for approximate computing, and software/hardware for neural network computation. Initial work, ongoing at the start of the NSTRF grant, involved a hardware implementation of an optical flow algorithm, used for robotic navigation in Field Programmable Gate Array (FPGA) hardware. This algorithm, unfortunately, was not competitive with state of the art approaches that were not biologically-inspired. Subsequent work focused on more general hardware originating from the postulate that artificial neural network hardware can provide general benefit to computing systems, including robotics. This benefit stems from the fact that dedicated hardware provides multiple order of magnitude energy-efficiency improvements over software. Towards this goal, I developed fixed-topology1 neural network accelerator hardware and demonstrated and used it to approximate transcendental mathematical functions in shared libraries providing two orders of magnitude energy-efficiency improvements. The approximation of shared libraries had not be done before and provided a low-overhead technique to approximate computing that avoided the difficulties of program analysis. Finally, and forming the majority of my PhD thesis, I developed a set of software and hardware extensions, X-FILES, that expose neural network computation as a first class functional primitive. I then developed a neural network accelerator backend for the X-FILES called DANA. This combined work, X-FILES/DANA, was integrated with a RISC-V microprocessor and demonstrated performance improvements of two orders of magnitude when evaluated on an FPGA. While neural network accelerators have been developed for some time, this work takes a system level view, encompassing hardware, user, and supervisor software, to truly expose neural network computation as a generic functional primitive to all applications running in a system.