Edge Inference for Robots and Satellites

Motivation

Effective and efficient operational deployment of machine learning inference algorithms poses many challenges. Especially in applications like mobile robots and satellites a high degree of autonomy can be achieved by using inference for decision-making. However, particularly in these resource constrained conditions, inference optimization plays a key role to use these limited resources as efficiently as possible. Here, optimization can target one or multiple metrics: Minimize power consumption, maximize inference speed or minimize computing resource usage.

Additionally, modern small satellites can carry high resolution and multispectral image sensors or other instruments generating large amounts of data. This poses additional challenges: Large amounts of data are challenging to store aboard small satellites and cannot be downlinked to the base station quickly. Here, machine learning algorithms can be employed to pre-select data and discard faulty or uninteresting sets or images.

To accelerate these operations, we investigate how to use the limited resources aboard robots and small satellites as efficiently as possible when operationally deploying machine learning algorithms.

Modern Payload computers

Satellite onboard computers as well as dedicated payload processing computers for small satellites are using more and more commercial-off-the-shelf (COTS) processors or whole single-board-computers (SBC). To reduce cost and shorten development times, components like different versions of the Raspberry Pi or Nvidia Jetsons are being used to process the large amounts of data generated by modern satellites.

These energy-efficient and - compared to previous generations of small satellite computers - very powerful processing solutions enable the deployment of state-of-the-art machine learning models for onboard processing.

Current research topics

• Analysis of inference optimization and hardware architecture
  • Raspberry Pi 4 / 5 (ARM); x86-Laptop class processors
  • Influence of compiler optimization
  • Influence of Quantization
• Development of FPGA SOC payload computer with EM-Munich
  • Smart Processing and Power Module (SPPM): https://www.em-munich.space/
  • Planned integration into University of Würzburg Experimental Satellite 5 (UWE-5) with launch ~2027