Acceleration Robotics Promises Big Performance Gains for ROS 2 Robots via Its ROBOTCORE Framework

Acceleration Robotics has lived up to its name with the launch of ROBOTCORE, a framework it claims can leverage FPGAs and GPUs alike to help boost the performance of Robot Operating System 2 (ROS 2) devices considerably — making them up to 500 times faster, in some cases.

"Robots are networks of networks, with sensors passing data to compute technologies and actuators. These networks can be understood as the nervous system of the robot," explains Víctor Mayoral-Vilches, founder of Acceleration Robotics. "Like with the human nervous system, low latency and real-time information is fundamental for the robot to behave coherently.

"Faster robots (or with more dexterity) require faster computations," Mayoral-Vilches continues. "Hardware acceleration with ROBOTCORE empowers exactly this. With ROS being the common language roboticists use to build 'robot brains,' ROBOTCORE extends ROS and deals with GPU and FPGA vendor-proprietary libraries, empowering hardware acceleration across silicon vendors."

That latter is key to the ROBOTCORE story: While hardware acceleration, either by using custom gateware running on an FPGA or by offloading highly-parallel workloads to a GPU, is well established in robotics, Acceleration Robotics is hoping to make things easier on developers by providing a framework for the development of custom compute architectures that remain entirely agnostic as to the hardware on which they're running — both on the robot and the accelerator side of things.

At launch, ROBOTCORE supports 12 host devices: The K26, KR260, KV260, ZCU102, and ZCU104 from AMD; the Jetson Nano, Jetson Nano 2GB, Jetson Xavier NX, Jetson AGX Xavier, and Jetson TX1 from NVIDIA; the PolarFire Icicle Kit from Microchip, the subject of our ongoing FPGAdventures series; and the Ultra96-V2 from Avnet.

The precise gains available depend on both device and workload: In the company's testing, ROBOTCORE running on an AMD KV260 showed a 3.96× increase in performance-per-watt; the ROBOTCORE Perception add-on boosted operations from 2.61× for resize up to 509.25× for the generation of a histogram of oriented gradients and boosted three-node perception graph pre-processing performance-per-watt 4.5×; while the off-device ROBOTCORE Cloud add-on delivered a fourfold reduction in ORB-SLAM2 simultaneous localization and mapping (SLAM) runtime and a 28.9× reduction in motion planning templates (MPT) compute runtime, including the time it took to upload the data and download the result.

More information on ROBOTCORE and its Perception, Transform, and Cloud extensions are available on the Acceleration Robotics website; pricing, however, is available only on application.