Ubotica CogniSAT-XE1

AI and Computer Vision
Edge Computing Platform Overview

The power of EdgeAI brought to small satellites

The Ubotica CogniSAT-XE1 On-Board AI Payload Coprocessor brings the power of Computer Vision (CV) and Artificial Intelligence (AI) compute acceleration to a PC/104 form-factor for SmallSat and CubeSat missions. It is built around the Intel® Movidius™ Myriad™ 2 CV and AI COTS Vision Processing Unit (VPU) whose 12 vector cores provide high-performance parallel and hardware accelerated compute within a low power envelope.

Widely Adaptable for AI and Computer Vision Deployments

Flexible and Integration Ready

CogniSAT-XE1 combines the power efficient compute of the Myriad 2 VPU with a wide range of interfaces and peripherals, providing broad flexibility for integration into satellite platforms. Either Gigabit Ethernet or USB2.0/3.0 can be used as the primary control and data interface to the board, enabling data rates sufficient to handle many CV and AI applications at near-streaming throughput.

Efficient Neural Network Inference

Common Neural Network (NN) frameworks (e.g., TensorFlow, PyTorch, Caffe) can be used for NN model development and training, with the model subsequently compiled to target the Myriad 2 device using Intel®’s OpenVINO™ toolkit. Workflows contained within the CogniSAT-TK package guide this process to reduce error. CogniSAT-XE1 leverages the broad range of pre-qualified models and layers available within OpenVINO™. Pre-trained OpenVINO™ models can be used with transfer learning, or can be deployed directly on the CogniSAT-XE1 board using the Host Control Software (HCS) which is supplied as part of the package.

A host loads NN models to the CogniSAT-XE1 board, subsequently sending image frames and receiving processing results. The host can dynamically update the CogniSAT application over the primary interface, enabling flexible runtime solutions.

Enabling an Ecosystem of features

Ubotica supplies a range of utilities, as part of the CogniSAT-TK package, which allow the HCS supplied with CogniSAT-XE1 to operate seamlessly with most sensors currently on the market. Ubotica and our ecosystem partners are constantly expanding the set of utilities as new sensors and functions are identified.

Deployment to CogniSAT-XE1 involves the transfer of only a single configuration file, and runtime updates enable the updating of pipelines without requiring re-compiles or system reboots.

Key Features

CogniSAT-XE1 has the
following features:

Part of the CogniSAT Edge Computing Platform
PC/104 form factor
Flight Proven Heritage
On-board integrated latch-up protection, with auto-restart and trip signal
Single 5V supply
CAN interface for C&C
CV & AI engine operating as a slave over USB 2.0/3.0 or Ethernet

Serial NOR boot flash
External power-down control
Parallel and MIPI camera interfaces for direct sensor connection
Fully compatible with the Ubotica CogniSAT Software, for complete control over AI processing
microSD card for non-volatile data storage

Additional on-board peripherals
Floating Gate Dosimeter
IMU (3D accelerometer and gyroscope)
Operates from single 5W power supply
Stackable 2x dual row board-edge headers for power delivery and low-level control

CogniSAT-TK™

Image Processing Toolkit

Ubotica’s CogniSAT-TK is a lightweight library for developers to bring the power of Myriad’s Computer Vision (CV) and Artificial Intelligence (AI) compute acceleration to their in-orbit applications.

When combined with other elements of the CogniSAT family, the CogniSAT-TK allows the full feature set of the Intel® Movidius™ Myriad™ Vision Processing Unit (VPU) to be utilized to facilitate the rapid development and deployment of applications to process data on-board satellite in a fast, power efficient manner.

KEY FEATURES

Convert and compile Neural Networks using OpenVINO™
Integrate compiled networks with custom image processing for target hardware
Tile large images for more efficient image processing
Convert sensor-proprietary formats to a common structure
Normalise image intensity to aid in Neural Network performance

CogniSAT-HCS™

Host Control Software

Ubotica’s CogniSAT-HCS is a lightweight host control application which allows the execution of AI applications to be integrated within the Payload Processor software stack without any development load.

The selection of neural network to be run and images to be processed is handled by simply editing a JSON configuration file to point to the location of the Neural Network file and images to be processed.

Once this is done, the full power of the CogniSAT processing platform can be harnessed by the execution of a single line of code.

KEY FEATURES

Fully compatible with Myriad Neural Networks developed for Intel’s OpenVINO platform
Runtime configuration of Myriad’s Hardware and Software CV filters for Image Transformations
Integrated Neural Network Supervisor
Optimised for environments with a single Myriad VPU
Supports both USB and Ethernet connectivity
APIs for C/C++ and Python programming languages