Mobileye Global Inc. · NASDAQ:MBLY
Mobileye Global — CES 2022
53 slides from Mobileye Global's conference presentation, 2022
Slide 1: Mobileye Under the Hood
Mobileye Under the Hood Prof. Amnon Shashua President & CEO CES 2022 © 2022 Mobileye
Slide 2: 2021 in Numbers
2021 in Numbers EyeQ Shipped 28.1 41 50M 30 24 New design wins with Pipeline volume of new 19.3 30+ OEMs design wins (compared 17.5 Record figure 18 to 37M in 2020) 12.4 8.7 12 188 1.4B 6 2.7 4.4 6 Vehicle models Revenue in 2021 launched in 2021 40% YoY 2014 2015 2016 2017 2018 2019 2020 2021 with Mo
Slide 3: The Mobileye Database
The Mobileye Database Believed to be the largest driving database in the industry Data: 200 PB 16M CLIPS premise-On + cloud (AWS) 25 years of driving Reference numbers: Intel (238), Israeli governmental services (5), Israel’s biggest insurance company (2) Compute: 500K 50M PEAK CPU MONTHLY RUNS CORE
Slide 4: Mobileye Strategy
Mobileye Strategy 2017 2021 2025 The fundamentals of Mobileye’s strategy, as we outlined back in 2017: REM™ MAPPING TRUE REDUNDANCY™ RSS SAFETY MODEL + Crowdsourced AV-maps + Computer vision that powers the + Formal safety guarantees + Cloud-based enhancements for ADAS Sense-Plan-Act cycle end-to-en
Slide 5: Mobileye Strategy
Mobileye Strategy 2017 2021 2025 True Redundancy, REM, and RSS are the building blocks for: LEVEL 2+ LEVEL 3 / 4 Premium ADAS Conditional autonomy / full self-driving + Driven by camera-only subsystem- full ODD + True Redundancy paves the way for high MTBF at low cost + REM enables scale + REM- clou
Slide 6: Achievements
Achievements 2017 2021 2025 Where we are today: REM™ MAPPING TRUE REDUNDANCY™ RSS SAFETY MODEL + We built the largest crowdsourcing fleet for + SuperVision™- productizing CV subsystem for + IEEE 2846 working group mapping- 25M km collected daily hands-free L2+ + Chaired by Intel-Mobileye with 30+ le
Slide 7: The Largest Global Footprint in the AV Industry
The Largest Global Footprint in the AV Industry ISRAEL DETROIT NYC TOKYO MUNICH PARIS 7
Slide 8: AV driving in Paris
= PRND AV driving in Paris Chon 8 iMmobileve | BYP ‘ a Bde moumey " (i ft i AA =f BW /
Slide 9: AV driving in Tokyo
= t AV driving in Tokyo | i COMFORT 25 50 2s Hit t }
Slide 10: Achievements
Achievements 2017 2021 2025 Where we are today: CLOUD-ENHANCED L2 LEVEL 3 LEVEL 4 Honda Legend, Japan only Volkswagen Travel Assist™ 2.5 + Unveiled Robotaxi at IAA, on the road 2022 (computer vision) (powered by REM maps) + Signed Robotaxi/ AV shuttle/ goods delivery deals Next-generation Ford Blue
Slide 11: First Design Win Consumer L4 Platform
First Design Win Consumer L4 Platform Powered by 6 EyeQ®5 High SoP early 2024 11
Slide 12: Where Does This Lead Us?
Where Does This Lead Us? 2017 2021 2025 New category emerging: Emerging 2022/2023: Emerging 2024/2025: L2+ Premium ADAS L4 Robotaxi L4 Consumer AV + Surround sensing + Geo-fenced + Drives everywhere + Cloud-based enhancements- REM and OTA + Cost- $X0,000 + Cost- <$5K + Full Sense-Plan-Act cycle on a
Slide 13: Where Does This Lead Us?
Where Does This Lead Us? Robotaxi Waymo, Argo, Cruise, Aurora, etc. The autonomous (people and goods) mobility market may unfold in two forms Consumer AV Tesla, Apple? Consumer AV Robotaxi PTO/TNCs buy AVs and just add the service layer on top Robotaxi Consumer AV Geo scalability and consumer-level
Slide 14: The Criteria for a Good Solution
The Criteria for a Good Solution Capabilties Robustness Efficiency Wide self-driving ODD Cost (compute and sensors <$5K) Definition Human-like driving policy High MTBF Scale (drive everywhere) Mobileye’s + Full ODD from L2+ to L4 + RSS (formal safety) + Purpose-built SoC approach + What differentiat
Slide 15: Purpose-built SoC, SW-defined imaging radar, Lean Compute
+ Purpose-built SoC + SW-defined imaging radar + Lean Compute + REM crowdsource mapping 15
Slide 16: The New Generation of EyeQ
The New Generation of EyeQ® 16
Slide 17: A Family of SoCs That Covers the Entire ADAS/AV Spectrum
A Family of SoCs That Covers the Entire ADAS/ AV Spectrum The New Generation of EyeQ® EyeQ®6 Light EyeQ®6 High EyeQ®ULTRA L1-L2 L2+ / L3 L4 ADAS Premium ADAS Full Self-Driving 17
Slide 18: The EyeQ ULTRA
The EyeQ® ULTRA AV-on-Chip: A single SoC to power autonomous driving end-to-end + Controlling the entire AV stack allows us to know precisely what is required from the AV’s onboard compute + We first built an AV and only then designed an application-specific SoC for the AV + EyeQ® ULTRA utilizes an
Slide 19: EyeQ 6 High
EyeQ® 6 High The ultimate compute platform for Premium ADAS + 3x more TOPS than an EyeQ5H with just 25% more power + Advanced visualization capabilities for parking and UX applications supported by dedicated GPU, ISP, and a video enc. + Will carry all premium ADAS tasks (next-gen SuperVision™) 7 34
Slide 20: EyeQ 6 Light
EyeQ® 6 Light The new work horse for all ADAS functions + The most power/performance/cost-efficient SoC we have ever built + Compared to EyeQ4M: + 45% smaller package + 450% more TOPS + Similar power consumption + One-box, behind windshield solution that can support CV+ localization to power L2+ fun
Slide 21: Redefining Radars and LiDARs
Redefining Radars and LiDARs + Purpose-built SoC + SW-defined imaging radar + Lean Compute + REM crowdsource mapping 21
Slide 22: The Motivation Behind LiDAR and Radar Development
The Motivation Behind LiDAR and Radar Development 2022 LiDAR/radar subsystem 2025 LiDAR/radar subsystem • ToF LiDAR - 360° coverage • Front sector- 3-way redundancy • Remaining FoV- 2-way redundancy of cameras + radars • Advanced stock radars- 360° coverage • Massive cost reduction that will unlock
Slide 23: SW-Defined Imaging Radar - Game-Changing Capabilities
SW-Defined Imaging Radar - Game-Changing Capabilities Main attributes 4.0cm + Massive MIMO radar with 2,304 virtual channels (48x48) + High angular and vertical res.- 0.5°x2° + 100dB dynamic range + 40dBc azimuth sidelobe levels (SLL) + 20 FPS, Multi-mode scanning (SRR, MRR, steerable LRR) 14.5cm +
Slide 24: Separating objects in dense traffic
Separating objects in dense traffic Construction area and path delimiters detection 24
Slide 25: Object detection in a noisy environment
Object detection in a noisy environment (metal fence, tunnel) 25
Slide 26: Tracking a motorcycle in dense traffic off-boresight
Tracking a motorcycle in dense traffic off-boresight 26
Slide 27: Can EyeC Radar output be presented like LiDAR?
RESEARCH THEORY 1: Can EyeC Radar output be presented like LiDAR? RADAR LiDAR 27
Slide 28: Can EyeC Radar output be presented like LiDAR?
RESEARCH THEORY 1: Can EyeC Radar output be presented like LiDAR? LiDAR LiDAR RADAR! 28
Slide 29: LiDAR RADAR! LiDAR
LiDAR RADAR! LiDAR RADAR! 29
Slide 30: Can we train a network to create camera-like video from radar?
RESEARCH THEORY 2: Can we train a network to create camera-like video based on our Imaging Radar point Cloud? 30
Slide 31: Building the Best-in-Class FMCW LiDAR
Building the Best-in-Class FMCW LiDAR 16cm Main attributes 8.5cm + Best point density (600pt/deg^2, over 2M discrete 4D PPS over 1000 lines/sec) + Doppler content provides objects velocity and heading GoldBox- optical head enabling without dependencies on multi-frame tracking/registration 90 vertica
Slide 32: What We Have Covered So Far
What We Have Covered So Far Game-changing developments going forward: AV-on-Chip SW-defined imaging radar becomes a Front FMCW LiDAR, yielding tri-fold heavy lifter in the self-driving architecture redundancy in the front sector 32
Slide 33: Lean Compute enabled by RSS Driving Policy Methodology
Lean Compute enabled by RSS Driving Policy Methodology + Purpose-built SoC + SW-defined imaging radar + Lean Compute + REM crowdsource mapping 33
Slide 34: Sense / Plan / Act Methodology
Sense / Plan / Act Methodology Sense Plan (Driving Policy) Act Perception of the environment. Decision making Execute the plan (Control): transform speed and curvature Building a world model of the vehicle's “What would happen if" commands to pedals and steering surroundings: where we are, other roa
Slide 35: About Driving Policy
About Driving Policy Definition: Sensing state Action The sensing state contains our location, The desired action static obstacles, lane semantics, traffic lights, is the speed and kinematic state of other road users, etc. curvature of the car Why it is hard? + No “Ground Truth” + Actions may have a
Slide 36: Driving Policy — Existing Approaches
Driving Policy — Existing Approaches Compute X Brute force Monte-Carlo Tree Search (MCTS) X Markov Decision Process (MDP) Quality of search MDP MCTS Brute force Realism of assumptions X X X Assumes the Assumes the No assumptions future behavior of policy of all road all road users is users is known
Slide 37: Our Approach - RSS
Our Approach - RSS A formal model for safety, The Method that provides mathematical formalization for the AV’s driving policy to never cause an accident 01 Defining reasonable boundaries on the behavior of other road users http://arxiv.org/abs/1708.06374 02 Within the boundaries specified by RSS, on
Slide 38: RSS Standardization and Government Efforts
RSS Standardization and Government Efforts ISO/TR 4804:2020 IEEE WORKGROUP TO DEFINE A FORMAL MODEL SAFETY AND CYBERSECURITY FOR FOR AV SAFETY CHAIRED BY INTEL-MOBILEYE AUTOMATED DRIVING SYSTEMS • Workgroup consists of 30 leading industry players • World’s first ISO Technical Specification defining a
Slide 39: Our Approach - RSS
Our Approach - RSS By using induction and analytical calculations, the RSS couples all plausible futures into the present This yields efficiency, realism, quality, and explainability Instead of “Predictions” we Construct “intentions” of other agents • Those “intentions” control parameters of the “re
Slide 40: Our Approach - RSS
Our Approach - RSS Compute X Brute force Monte-Carlo Tree Search (MCTS) X Markov Decision Process (MDP) RSS-based driving policy Quality of search MDP MCTS Brute force Realism of assumptions X X X RSS-based driving policy 40
Slide 41: More on Mobileye's Lean Driving Policy
More on Mobileye’s Lean Driving Policy From our CTO, Prof. Shai Shalev-Shwartz Search: “Mobileye’s Lean Driving Policy” 41
Slide 42: Advancements in REM Mapping
Advancements in REM Mapping + Purpose-built SoC + SW-defined imaging radar + Lean Compute + REM crowdsource mapping 42
Slide 43: Volkswagen Travel Assist 2.5
Volkswagen Travel Assist 2.5 The first L2+ system to widely use Mobileye’s REM technology 43
Slide 44: Volkswagen Travel Assist 2.5
Volkswagen Cloud-enhanced lane-centering in challenging scenarios Travel Assist 2.5 The first L2+ system to widely use Mobileye’s REM technology 44
Slide 45: Traffic light-to-lane association based on REM map
Traffic light-to-lane association based on REM map 45
Slide 46: Europe REM RoadBook Coverage
Europe REM RoadBook Coverage + 2.5 million km + Generated fully automatically in the cloud in less than a week 46
Slide 47: km harvested in 2021
km harvested in 2021 1750 M 1541 M 470 M 247 M Q1 Q2 Q3 Q4 47
Slide 48: The Richness of REM AV Maps
The Richness of REM AV Maps Main attributes of REM AV maps provided in any road type, as we revealed last CES: + Drivable paths + Road edge + Traffic light and Traffic sign to lane association + Yield and priority + Crosswalks and crosswalks relevancy + Stopping points and stop lines + Common speed
Slide 49: Advancements in REM Development
Advancements in REM Development In 2021, we have added new features to REM maps, all based on crowd knowledge: Crowd Turn Indicators Construction Area Live Map Determining where to apply the Turn indicators through crowdsourcing for true “Human-like” behaviour of the AV In 12 out of 15 samples, turn
Slide 50: Advancements in REM Development
Advancements in REM Development In 2021, we have added new features to REM maps, all based on crowd knowledge: Legal speed Speed bumps Lane Level Legal Speed Limit Indication Enables smooth control in urban and rural areas Accounting for Explicit Speed Limit Signs, Implicit Speed Limit Signs and Roa
Slide 51: Advancements in REM Development
Advancements in REM Development Enriching the map by adding Semantic Lane Types : Road type Toll areas Public transportation lanes Refining policy parameters accordingly Also relevant for L2+ applications (e.g. pedestrian on a highway vs. deep urban) 51
Slide 52: Summary
Summary The building blocks we have built: 01 Redefine the future of ADAS with REM mapping and CV subsystem 02 The right engineering design to achieve the needed MTBF and unlock MaaS 03 AV-on-chip and SW-defined imaging radar redefine the future of consumer L4 52
Slide 53: Thank You
Thank You Mobileye Under the Hood Prof. Amnon Shashua President & CEO CES 2022 © 2022 Mobileye