XCube Accelerating Autonomous Driving Development February 2017 XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 1 To the Winner go the Spoils • Autonomous driving (AD) is a high-stakes race between – Disrupters (Waymo (Google), Tesla, NuTonomy, Uber) and – Incumbents (GM, Toyota, et al and Tier 1 and 2 suppliers) • The same skills are needed for ADAS • At stake are trillions of dollars in market share* That’s why: – GM bought Cruise for $1 billion in March 2016 – Ford bought Argo for $1 billion in February 2017 *ADAS will be $132 billion a year by 2026. ABI Research, 2016. XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 2 Advanced Driver Assistance Systems (ADAS) Autonomous Driving Curve speed adaptation Collision avoidance Lane departure warnings Automated braking Parking Automation Blind spot warnings © 2017 XCube - All Rights Reserved XCube Confidential Information 3 Autonomous Driving (AD) and ADAS both require collecting massive datasets Maps Software Data Sets Information Modification Sensor Data Decision Making Cameras LIDAR XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 4 Developing AD or ADAS software - simplified Images Maps Sensor Data Data Pool Rules Decision Result Application Deep Learning / Neural Networks XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 5 The Players XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 6 Core Challenges… Computer 1st Trainingtothe Drive 2nd Proving AD/ADAS Safe XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 7 Training Challenges Need millions of test drive scenes and variations How can one find specific scenes in the petabytes of data in multiple locations worldwide? XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 8 Logic needs to be simulated under many conditions These Include: • Road condition and type • Traffic conditions • Weather conditions • Lighting conditions • Object interactions – Ball in the road – Traffic signals • Other vehicle behaviors – Brake lights two cars ahead would cause front car to break XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 9 Why is it difficult? • Current tools are inadequate – Hadoop and Spark require expensive and time consuming centralization of data and a new application. Even if you are using newer data lakes, however: • Deep learning algorithms incapable – Of knowing what to test – Of testing and reasoning about particular scenes – Of being scaled in parallel – Tedious to re-test similar scenes “Machine learning systems, and deep learning in particular, are very high performance but don’t actually explain how they figure out the answer” – Gill Pratt, CEO Toyota Research Insight XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 10 In sum, developing AD or ADAS software requires: • Finding unusual situations and catastrophes in petabytes of data • Simulating the unusual situations • Acceleration of simulation in parallel • Training deep learning neural networks in parallel on specific scenes XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 11 XCube Toolset is Solving Those Challenges Main software functions Store • Store hundreds of petabytes on geographically distributed systems Mine • Tag content inside the data files automatically Search Use • Search the tagged content in parallel for matching time intervals • Run any desktop application (e.g. simulator) in parallel Manage • Manage millions of search results and billions of test results Automate • Automate regression testing, AI-training, and proof of safe operation XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 12 XCube Automates Machine Learning at Scale • Transforms problem of distributed data into the advantage of parallel processing • Automates training and testing of neural networks in parallel • Allows granular testing of situations, such as rain, snow, or four-way intersections • Vendor independent framework support – Supports most popular frameworks natively (Caffe, Torch, Tensor Flow, Matlab) – NVidia, Intel, Xeon-Phi, new FPGA-CPU and other architectures XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 13 Team Satish Jha, Chairman & CEO Mikael Taveniku PhD, President • CIO, Roche Holding AG • CEO, James Martin & Company • Seeded and mentored 35 startups • EDHEC; Fletcher School of Law and Diplomacy 27 years of leadership in: • Ericsson, Mercury Systems • High-end designs of EASA radar, sonar, CT • Machine learning advanced applications • PhD, MECS, BSEE, BSME Yaron Naor, VP Business Development Shantanu Jha, Chief Product Officer Founder & CEO, Mobile Wisdom, a ComverseSiemens JV GM at Comverse Commander (Ret.) Israeli Defense Force MBA, University of Aalto • Portfolio Manager, Beaconsfield Capital Management • CTO, Uni AI Software • 10 years of deep learning experience • University of Chicago, LSE, Cambridge • • • • © 2017 XCube - All Rights Reserved XCube Confidential Information 14 XCube Initial Customers - Ongoing PoCs XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 15 Current Use Cases • Train the deep learning system for all situations • Prove that the system is responsive, comprehensive and safe • Find particular driving scenes from millions of hour of recordings • Automatically annotate data from video and other sensor streams • Run new algorithms in parallel on tens of thousands of scenarios found, varying hundreds of parameters in each scene • New advanced algorithms with deep learning under development • Data and team distributed across the world • Run applications based on distributed datasets from anywhere • Very large scale simulations XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 16 XCube Pipeline OEMs Tier 1 Suppliers Tier 2 Suppliers XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 17 Revenue growth – One Customer, One vehicle Model $100K 3 – 9 month Pilot Program $1M 1st year 3 month initial deployment Year 1:Small Team $5M / year Vehicle testing on road Year 2-4: Global team Fleet test, verification 15% - 20% Annual Maintenance Model on Road ~Years 3 – 10 $15M - $20M Revenue Per Vehicle Model XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 18 Financial projections Today Pro Forma Income Statement ($M) 2016 2017 2018 2019 2020 (actual) Revenue 1.1 3.5 12 27 55 COGS 0.6 1.7 5.0 12 25 Margin 0.5 1.8 8.0 15 30 R&D 0.3 1.0 3.5 7.0 10 SG&A 0.1 0.7 2.0 3.0 5.0 Net Profit 0.1 0.1 2.5 5.0 15 XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 19 Milestones Secure funding and hire additional engineers / team members Feb 2017 June 2017 Secure at least 3 major deals with OEMs and/or Tier 1s Q1 2018 Dec 2017 Complete product evaluation with Delphi, Honda, SL Corp and Baidu Win additional customers and establish office in APAC Q2 2018 Q3 2018 Expand local presence in Europe XCube © 2017 XCube - All Rights Reserved XCube Confidential Information 20 2 – 3 Year Exit Strategy • Candidates – Potential Users (SAIC, Hyundai) – Prospects (Delphi, Volvo, Toyota, GM) – Technology Partners (ÅF, NVidia) • Comparable Exits Cruise Automation ($1B, no revenue) Automotive: Argo ($1B, no revenue) Actuate ($330M) Big Data: Machine Learning: Pentaho ($550M) ColdLight ($105M, $8M in revenue) AlchemyAPIàIBM WhetlabàTwitter XCube © 2017 XCube - All Rights Reserved * Source JD Ford & Company XCube Confidential Information 21 Summary Marquee Customers – Honda, Volvo, Delphi, Baidu, SAIC, Huayu Unique Product – State of the art product automating machine learning/testing for autonomous vehicles Supportive Investors - Beacon Angels, Launchpad Organic Growth Business Model – becoming a long term partner providing scalable infrastructure Stable Team – top tier in technology, management, and business development © 2017 XCube - All Rights Reserved XCube Confidential Information Thank You © 2017 XCube - All Rights Reserved XCube Confidential Information 23
© Copyright 2026 Paperzz