A group of researchers from MIT and Harvard are looking to address a key issue in the world of robotics: the gap between how quickly robots can process their surroundings and respond to them.
Generally, complex robots operate in three steps. First, it perceives its surroundings, gathering data using sensors or cameras. Then, based on what it perceives, it maps out the world around it and localizes itself within that map. Finally, it plots its course of action according to what it maps out, a process known as “motion planning.” Doing all these steps correctly and acting on them takes a lot of time and computing power.
However, using a method called “robomorphic computing,” the researchers designed custom computer chips that help robots produce faster response times — a process known as hardware acceleration. An extended abstract of this research was recently made publicly available, and the full paper will be presented at the International Conference on Architectural Support for Programming Languages and Operating Systems in April.
In a nutshell, these hyper-specific chips help with motion planning by taking into account a robot’s surroundings and intended use. Specific parameters are then set according to a robot’s physical form (its limb topology or joint types, for example), which then creates an accelerator customized to that specific robot model.
The paper says that, until now, relatively little work in hardware acceleration has been done for motion planning. Instead, the focus has been on areas like perception and localization. And the work that has been done in motion planning has largely been centered around the problem of collision detection for systems with “simple” dynamics like cars and drones. This latest hardware, however, can be applied to a “wide variety of complex robots.”
A classic example of hardware acceleration is a graphics processing unit, which was explicitly designed to accelerate the creation of images that display on a computer. When compared to the hardware designed using robomorphic computing, the custom chip outperformed off-the-shelf CPUs and GPUs. Although it operated at a slower clock rate, the chip performed eight times faster than the CPU and 86 times faster than the GPU.
Sabrina Neuman, a recent Ph.D. graduate from the MIT Computer Science and Artificial Intelligence Laboratory who led this research, says she was “thrilled” with these results.
“Even though we were hamstrung by the lower clock speed, we made up for it by just being more efficient,” Neuman told MIT News, which first reported the story.
Broadly speaking, the researchers believe that the burgeoning area of robomorphic computing provides a “reliable shortcut” to the otherwise tedious and error-prone traditional hardware accelerator design process. It could also help robots better take over more risky tasks for humans, like caring for patients in a hospital or moving heavy objects.
“[Our methodology] represents meaningful progress toward real-time, online motion planning and control for complex robots, the performance of which is limited by current software solutions,” the paper concludes. “Using robomorphic computing to shrink this performance gap will allow robots to plan further into the future, helping them safely interact with people in dynamic, unstructured, and unpredictable enigmatic environments.”
As far as robomorphic computing’s potential, Brian Plancher, a researcher at Harvard who co-authored the paper, says this is just the beginning.
“Ideally we can eventually fabricate a custom motion-planning chip for every robot,” Plancher told MIT News. “I wouldn’t be surprised if, 20 years from now, every robot had a handful of custom computer chips powering it, and this could be one of them.”
