There’s been a ton of innovation around drones over the past few years. Several companies are already testing drones for package and food deliveries, along with using drones for other use cases like agricultural monitoring, inspecting critical infrastructure such as power lines, inspecting fire damage, and more. It’s clear that there’s a huge demand for both remote-controlled and autonomous drones in the commercial sector, and that demand will only ramp up further as regulations for drones are sorted out. This article will explore how the latest AI computing advancements are ushering in a new era of innovation for drones.
One of the critical challenges for drones is that they need extremely powerful processing that is also incredibly energy-efficient and in a small form factor. This is doubly true for drones that use high-definition or multiple cameras for computer vision (CV) applications. Today’s digital computing solutions are often power-hungry, impacting the flight times and battery life of drones. Furthermore, digital computing solutions have trouble running complex AI networks, which are critical to providing immediate and relevant information to the control station. Just imagine a drone used to monitor an oil rig for leaks or other signs of damage. The drone needs to process the footage instantly to detect problems in real-time and immediately report them.
While digital solutions aren’t sufficient for the challenging requirements of many drone AI applications, a new computing approach can help eliminate these hurdles. Analog compute in-memory (CIM) can perform real-time AI processing – even with multiple large, complex deep neural networks (DNNs) – at a fraction of the power of a digital processing system. Analog CIM systems work by pairing analog compute with a non-volatile memory (NVM) like flash memory, unlike digital computing systems that rely on high-throughput DRAM. Whereas DRAM consumes a lot of power, analog CIM systems have significant power advantages by performing massively parallel vector-matrix multiplication and addition operations inside flash memory arrays.
Analog CIM systems also do not suffer from the latency of data propagating through digital logic gates and memory in the processor and written and readout of external DRAM, meaning that analog CIM systems can process compute-intensive AI workloads extremely quickly. Additionally, analog CIM systems are very compact, essential for drones with size and payload (i.e., weight) restrictions. Analog CIM systems can deliver big AI processing power in a small form factor thanks to the high flash density, making it possible to use a single flash transistor as a storage medium, multiplier, and adder (accumulator) circuit.
All of these factors make analog CIM systems ideal for a wide variety of AI video analytics applications for drones, including object detection, classification, segmentation, and depth estimation. These capabilities will open up new and exciting possibilities for drones in the coming years.
As drones can process more information locally, there will be a rise in drones that can operate entirely autonomously. These drones will be able to handle complex tasks for various industries, including agriculture, delivery, environmental protection, security, and more. Of course, there will still be plenty of applications where drones will need to be controlled by humans – whether for safety reasons, regulations, or other concerns – so increased autonomy will make it possible for one person to pilot multiple drones at a time instead of controlling just one.
In addition to the outdoor use cases for drones that are commonly discussed, there are also a number of ways that drones will be used indoors in factories and other industrial settings. Drones can help monitor and identify inventory and transport goods to different parts of the warehouse. Drones will also be important for inspecting equipment, especially in areas that are hazardous to people.
Finally, more capable drones will also drive advancements in counter-drone technology. The Federal Aviation Administration receives more than 100 reports a month about drones flying in restricted airspaces, such as airports. Even if these drones are being controlled by hobbyists who have no ill intentions, they can still pose a serious threat to aircraft, especially helicopters and small planes. Computing advancements will enable drones to get better at capturing other drones deemed a threat to public safety.
I look forward to seeing how powerful AI processing based on analog CIM technology will reshape the next generation of drones and open up new applications for almost every industry imaginable.