Speaker
Description
In the context of in-orbit servicing, the implementation of a space-servicing vehicle necessitates a specific level of autonomic decision-making capabilities. This requirement becomes particularly crucial when dealing with active debris removal of defunct satellites that exhibit uncooperative behaviour during the removal process, potentially also displaying unpredictable attitude behaviour. To address these challenges safely, computer vision plays a vital role in providing real-time visual perceptual and situational information to the lower-level control system of the space-servicing vehicle. To enable the reliable execution of computer vision tasks on board a spacecraft, Magics is currently developing a radiation-hardened AI processor specifically optimized for such computer vision workloads.
While several AI processors are entering the space market, it is worth noting that they are mostly upscreened commercial components not explicitly designed to withstand hostile environments like space. Given the safety-critical nature of in-orbit servicing tasks, it is crucial to ensure a high level of fault tolerance in the onboard computer. However, with upscreened components, this typically requires the use of triple-modular redundancy (TMR), which significantly increases the area requirements. Moreover, sustainability is a key concern, as the space-servicing vehicle needs to withstand long-term radiation exposure to avoid contributing to the space waste problem itself.
To address these challenges, a dedicated radiation-hardened AI processor is essential. Magics' MAG-AIA00101-SC is such a processor, designed to be radiation-hard by design, delivering high computational performance (10 TMACs/s) to support advanced AI workloads. It boasts a peak power consumption of only 1 W, making it suitable for small satellites with passive cooling requirements. The processor incorporates a programmable RISC-V microcontroller unit (MCU) and features a vector engine to facilitate real-time image pre-processing. Additionally, a large-scale dedicated AI engine is integrated to accelerate AI model inference. As the platform is optimized for computer vision tasks, it supports various modern peripherals essential for space missions with high data throughput demands. To further aid in AI model compilation and optimization, support for the open-source TVM platform will be provided.
In conclusion, the proposed radiation-hardened AI processor offers reliable and sustainable neural network inference capabilities for state-of-the-art computer vision workloads in space. This processor is particularly well suited for aiding control systems of satellites in safety-critical applications that require perceptual information at their input, as seen in active debris removal operations. It maintains competitiveness with commercially available components in terms of computational efficiency and software development ecosystem while providing the added advantage of being fault-tolerant and long-term reliable.