Using GreenWaves’ RISC-V to build the brain of nano-drone
“A few days ago, a research team published a paper demonstrating a parallel ultra-low-power (PULP) processor and convolutional neural network (CNN) that can give off-the-shelf Crazyflie 2.1 nanometer drones to achieve “top autonomous navigation capabilities” “――Although the size and weight are very small.
A few days ago, a research team published a paper demonstrating a parallel ultra-low-power (PULP) processor and convolutional neural network (CNN) that can give off-the-shelf Crazyflie 2.1 nanometer drones to achieve “top autonomous navigation capabilities” “――Although the size and weight are very small.
“AI-driven pocket-sized air robots have the potential to completely change the IoT ecosystem, acting as autonomous, inconspicuous and ubiquitous smart sensors.” The team claimed in the abstract of the paper. “With the dimensions of a few square centimeters, nano-sized unmanned aerial vehicles (UAV) are a natural adaptation for indoor human-computer interaction tasks, and the attitude estimation problems we solve in our work are also natural adaptations.”
“However, due to the limited payload and computing power of nano UAVs, the onboard brain can only use microcontrollers below 100mW. Our processors are based on the new parallel ultra-low power (PULP) architecture paradigm and deep neural network (DNN) ) At the intersection of the general development method of the visual pipeline, that is, covering various functions from perception to control.”
The team’s work is currently focused on the existing commercial parallel ultra-low power (PULP) processor, namely GreenWaves Technologies’ GAP8. The chip is designed for edge AI workloads, including nine processor cores, RV32IMC RISC-V instruction set architecture based on the free and open source RISC-V implementation, plus an extended XpulpV2 processor, increasing the hardware loop, Post-increment address for load and store operations, single instruction multiple data (SIMD) vector operators, and floating-point instructions for 8-bit and 16-bit data.
On this basis, the team ran a custom convolutional neural network (CNN) called PULP-Frontnet. Inspired by the Proximity network, PULP-Frontnet is customized for low-resource applications and is connected to the built-in QVGA grayscale camera in the target Crazyflie 2.1 nanometer drone. Although the entire system needs to be able to run on the GAP8 chip, the performance has proven impressive. Actual tests show that the network can run at 135 frames per second, power consumption is only 86mW, and memory usage is only 184kB, which enables the drone to have “top-of-the-line autonomous navigation capabilities”-even being able to calculate around the flight area The posture of the wandering human.
“[PULP-Frontnet帮助]The robot maintains a constant distance in front of them. “The team concluded.” Solving this kind of HDI on autonomous nano drones[人机交互]The problem is a challenging and valuable task in the field of Internet of Things. These robot assistants can be conceived as the next generation of ubiquitous IoT devices, ideal for human indoor operations. “