Robots and other “intelligent” devices need a lot of computational power for machine vision, path planning, manipulation, and other functions. One alternative to trying to put all that capacity onboard was to move some of it to the cloud, but new data management techniques and processors are enabling more efficient distribution of data across the edge, networks, and the cloud. Intel Corp. today announced four research papers, including one describing how a new accelerator could benefit artificial intelligence, robotics, and augmented reality.
Intel presented “A Ray-Casting Accelerator in 10nm CMOS for Efficient 3D Scene Reconstruction in Edge Robotics and Augmented Reality Applications” at the 2020 Symposia on VLSI Technology and Circuits. The organizers of the VLSI (Very Large-Scale Integration) virtual event are affiliated with the Institute of Electrical and Electronics Engineers (IEEE).
Edge robotics and augmented reality (AR) are among the applications that require rapid reconstruction of complex 3D scenes from large volumes of data for simultaneous localization and mapping (SLAM), said Intel. The company’s researchers devised a ray-casting hardware accelerator that they claimed can achieve both accuracy and energy efficiency.
Intel fellow discusses visual data accelerator
The techniques include voxel overlap search and hardware-assisted approximation of voxels, reducing demand on local access memory, according to Intel.
“A ray-casting accelerator in 10nm CMOS [complementary metal-oxide semiconductor] simultaneously casts multiple rays in spatial proximity to exploit voxel data locality, featuring a near-memory search for voxel address overlaps and opportunistic approximate trilinear interpolation for energy savings,” said Intel’s paper. “Measurements demonstrate ray-casting of 320×240 depth images with an average latency of 23.2ms/frame, while consuming 32.7pJ energy per ray-step and achieving a maximum energy-efficiency of 115.3 giga raysteps/W.”
“For computing voxel volumes, tens of millions of computations must happen, which must happen in the edge stack with processing constraints,” said Vivek De, Intel fellow and director of circuits research. “Our ray-casting hardware accelerator can improve energy efficiency by 30% without a loss in visual SLAM accuracy.”
“Our research started about two years ago in Intel Labs, which has a total of 700 researchers” he told The Robot Report. “A team of seven or eight researchers has been exploring different acceleration implementations and converged on one compelling approach to key challenges for data and intelligence at the edge.”
Basic research to benefit robotics
“Intel is conducting silicon building-block research to meet the demands of the most compelling dense visual SLAM applications,” said De. “It’s most applicable to when you want a detailed, accurate reconstruction of 3D data, such as in gaming, augmented reality, and some robots. Many robots use sparse SLAM, but this is more demanding and would fit in high-end robots that need to operate in complex environments.”
How long would it take for the ray-casting accelerator to be incorporated into commercial systems?
“We do the basic research, and Intel goes through fact-finding engagements with potential users and customers,” De replied. “In about three to five years, these building blocks will show up in the augmented reality and virtual reality.”
“Once we develop the accelerators, they would be standardized on SOCs [systems on a chip],” he said. “There won’t be much impact on the software stack, which can support accelerators of different types, such as video transcoding.”
“The accelerator could be useful for tele-operation or robotic surgery,” acknowledged De. “By providing more intelligent edge robotics without a loss in accuracy or energy efficiency, it could assist with higher-quality functions in a factory or a clinical setting.”