If I Only Had a Hardware Accelerated Brain

There is no disputing that the gap between what robots are capable of in science fiction and in real life is massive. It is hard not to be disappointed watching a robot slowly scan its environment, then jerkily inch along when what we really want is more along the lines of a responsive, fluidly moving, wisecracking C-3PO.

Modern robots are built with actuators that are entirely capable of fast, fluid, natural movements. When it comes to control, a major factor holding them back is processing power. After gathering sensor and camera data, a robot must first construct a map of its surroundings and localize itself within that map, then a motion plan needs to be calculated to achieve its objectives. These steps require a lot of computational horsepower.

CPUs are designed to be general purpose processors, and accordingly are good at a wide variety of computations, but are not necessarily optimal for any given task. This is where hardware acceleration can step in and speed things up. Certain tasks can be offloaded from the CPU, and onto purpose-built hardware that can perform the computations much more quickly. Probably the best known type of hardware accelerator is the graphics processing unit (GPU), which can process graphics data for thousands of pixels in parallel.

A team of researchers from MIT’s CSAIL have developed a new technique called robomorphic computing that helps develop hardware accelerated “brains” that are customized for a robot’s physical layout and intended applications. By customizing a processing chip for the specific use case, the response time of the robot can be lowered dramatically.

The system accepts parameters that specify characteristics such as the robot’s limb layout, and how the joints can move. Given this information, the system designs a hardware architecture that is optimized for the types of calculations that the robot will need to perform.

In a trial of the new method, a custom hardware design was tested with the help of an FPGA. Even though the clock speed of the FPGA was slower, it outperformed the tested CPU by 8 times, and ran 86 times faster than the tested GPU. They found that the efficiency of the custom design far outweighed the advantage of the higher clock speed of the off the shelf processors.

The researchers next plan to add some automation to the robomorphic computing method, which they hope will make designing custom hardware as simple as drag and drop.