Getting robots to move around safely in the physical world is tricky business. Industrial bots are big and powerful, and perfectly capable of crushing humans by accident, but kitting them out with robot vision and enough brains to avoid obstacles is costly and slows down movements. Usually, the bots simply operate on set paths instead, with humans warned to keep clear. New research from roboticists at Duke University could provide a practical solution to this problem, though — a new processor that can calculate where a robot should move approximately three orders of magnitude faster than current methods, all while using one-twentieth of the power.
As first reported by IEEE Spectrum, the chip in question is a customized FPGA or field-programmable gate array. As the name suggests, these are processors that can be reprogrammed after manufacturing to specialize in certain tasks. They’ve been around for decades, but are proving to be quite adept at problems involving machine learning. Microsoft, for example, is embracing FPGAs for its AI cloud services.
The advantage of using FGPAs for robots is clear. Usually when a robotic arm, for example, is shown an environment it needs to navigate, it takes several seconds to pause and calculate its route. It not only has to think about getting from A to B, but to calculate the 3D space it occupies in doing there. (This is known as the “swept volume” — think of it like the trail left by lights in long exposure photographs.) In the video above, you can see an FGPA-equipped robotic arm reacting almost instantaneously to new environments.
Speaking to IEEE, Daniel J. Sorin, a professor at Duke involved in the research said: “Motion-planning software has been a huge limiter to the adoption of robotics, and if you can do real-time motion planning, suddenly robots can now operate in dynamic, unstructured environments. That’s what we’re hoping to enable.”