ubotica technologies
play

Ubotica Technologies Space Activities Overview Providing AI - PowerPoint PPT Presentation

Ubotica Technologies Space Activities Overview Providing AI Solutions for Edge Based Computer Vision Applications Ubotica Introduction 2 Providing AI solutions for edge based Computer Vision applications Edge processing Process data at


  1. Ubotica Technologies Space Activities Overview Providing AI Solutions for Edge Based Computer Vision Applications

  2. Ubotica Introduction 2 Providing AI solutions for edge based Computer Vision applications • Edge processing • Process data at source rather than central node • “Photons in, metadata out” • Space is the ultimate edge! • Why process at the edge? • Power | Latency | Autonomy

  3. Ubotica Overview 3 Team of 15 split between Dublin and Madrid EMBEDDED AI & CV SPACE • Myriad MDK Development Licensee from Intel/Movidius • Preferred partner BIOMEDICAL AUTOMATION

  4. Ubotica Space Roadmap 4 Qualification HW SW System Credit: CERN/Maximilien Brice

  5. Myriad Vision Processing Unit 5 • Myriad 2 AI and CV engine 9.5mm x 8mm • 12 VLIW SHAVE vector cores • >1 TOPS • Vision hardware accelerators • Stacked DDR • Software reconfigurable hardware – software platform • 28nm HPC process node • Ideally suited for in-orbit • Low thermal budget • Low power budget (~1W) Credit: Intel Movidius

  6. Myriad Advantages 6 Neural Network Inference • Efficient embedded NN inference • 12 SHAVE vector engines and large cache-like CMX multi-ported memory • NN firmware optimised by Intel • Broad NN framework and network compatibility through OpenVINO • Caffe, TensorFlow, MXNet • Inception, ResNet, MobileNet, YOLO, SqueezeNet… • Large range of trained AI models • Models already qualified by Intel • Freely available and ready to use • Neural Compute Stick • Convenient development and hardware validation tool

  7. Myriad + Ubotica Advantages 7 Computer Vision (CV) and Image Signal Processing (ISP) • Efficient embedded CV and ISP functions on Myriad • Highly efficient hardware filters for common ISP functions • Library of optimised ISP and CV kernels executing in parallel on vector engines processors • Custom ISP/CV pipelines • Drag-and-drop Configuration Tool • Runs in-browser • Runtime reconfigurable • New pipelines can be submitted to Myriad at runtime • No re-compilation required • No interruption to embedded application

  8. Ubotica HW & SW 8 SW CVAI Toolkit for Myriad UB0100 CubeSat CV & AI Board HW HPCB EO Data Processing Board

  9. SW: CVAI Toolkit 9 • Unlock the CV and ISP capabilities of the Myriad • Enable users to develop their own applications • User API CV ISP • Utilise hardware blocks of Myriad 2 for efficient CV • Reconfigurable Image Signal Processing pipeline • Pre-process frames for inference • Perform optimised AI inference AI • Runtime interaction over Ethernet and USB • Integrates NN runtime dependability measure

  10. SW: CVAI Toolkit 10

  11. HW: UB0100 CubeSat Board for AI and CV Acceleration 11 • USB/Ethernet control and data interface, PC104 form factor • For more details see: https://www.ubotica.com/ub0100

  12. HW: HPCB Backplane Solution 12 • GR-VPX-XCKU060 and GR-HPCB-FMC-M2 High Performance Compute Board for EO data processing • Led by Cobham Gaisler • Specifications • FPGA • Rad-hard supervisor • Myriad 2s • Board available Q2 2021 • For more details see • https://www.gaisler.com/index.php/products/boards

  13. Myriad 2 Radiation Characterisation 13 • COTS device • Performance under cosmic radiation must be characterised • Highly penetrating ion beams required due to Myriad 2 SoC construction Credit: CERN/Maximilien Brice • Preliminary radiation characterisation • De-risk phase complete (low LETs) • SEL (CERN heavy-ion) • SEU (GSI heavy-ion) • TID (ESTEC Co-60) • No critical sensitivity observed • Further tests at higher LETs for full characterisation in 2021 • Heavy Ion, Proton/Neutron

  14. Ubotica Edge CVAI - Applications 14 • Ship identification & classification ( time critical ) • Hyperspectral filtering ( data reduction ) • Fire detection ( time critical ) • Navigation ( autonomy ) Credit: NASA

  15. Ubotica In-Orbit Use Case: PhiSat-1 15 • Detecting clouds in EO data in real-time • Demonstrate AI inference in-flight • Improve downlink bandwidth utilisation

  16. Ubotica In-Orbit Use Case: PhiSat-1 16 • Myriad 2-enabled in-orbit AI demonstrator • Flying on HyperScout-2 hyperspectral payload • Performing automatic cloud detection • Providing Myriad 2 in-flight performance data Ubotica AI Engine HyperScout-2 Credit: CERN/Maximilien Brice Credit: cosine

  17. Ubotica In-Orbit Use Case: PhiSat-1 17 PhiSat-1 Engineering Model Credit: Tyvak Ubotica AI Engine integrated on sensor payload Credit: ESA/Cosine

  18. Ubotica GNC Use Case: AIVIONICS 18 Credit: NASA • AIVIONICS project • AI for Guidance, Navigation & Control applications • Incorporating AI dependability scoring • Two use cases being studied • Moon/asteroid landing • Uncooperative object grasping (space junk collection) 22,000 trackable objects Credit: NASA

  19. Further Information 19 • Ubotica • Website • Article • Myriad 2 • Website • Product Brief • Technical Paper • Radiation Testing Media • ESA News Brief • High Performance Compute Board (HPCB) • Cobham Gaisler Press Release

  20. Ubotica Technologies Ltd. DCU Alpha Innovation Campus Old Finglas Road Glasnevin Dublin D11 KXN4 Ireland Providing AI Solutions for Edge Based Computer Vision Applications

Recommend


More recommend