0
Developing Autonomous Driving EVs for the
China Market XPENG Motors’ Approach
Junli Gu
VP of Autonomous Driving, XPENG
motors
March 18, 2019
1
Agenda
1. XPENG Motors Introduction and Background
2. Roadmap for XPENG Motors’ Autonomous Driving Solution
3. System Architecture Design Powered by Xavier
2
Who is XPENG?
We develop FUN, CONNECTED and INTELLIGENT EVs tailored for
the young and tech-savvy Chinese Millennials
2
• Aimed at young Internet users in China, with autonomous driving and
intelligent connection as the core differences
• 2014: Founded in Guangzhou, China
• Dec 2018: Mass production and official launch of G3 (electrical SUV)
• Q1 2019: Mass production of E28 (electrical sedan)
3
Strong R&D capabilities and investments
Shanghai
Intelligent Connected
Vehicle
Navigation R&D
Silicon Valley &
San Diego
Machine Learning
Autonomous Driving
R&D
Guangzhou
Vehicle and VCU
R&D
V2X Testing
Trial and Production
Lead-in
Beijing
Artificial Intelligence
R&D
R&D footprint
Our R&D headcount accounts for over 70% of our total number of employees.
4
Your Smart Driving Assistant
Connectivity
Xmart OS
X-pilot
Artificial
Intelligence
5
X-center
XPENG Headquarter
5
6
Agenda
1. XPENG Motors Introduction and Background
2. Roadmap for XPENG Motors’ Autonomous Driving Solution
3. System Architecture Design Powered by Xavier
7
L3 Products: Highly-Autonomous Highway Features
L4 Products: Fully-Autonomous
Urban & Highway Features
- Driverless Robo Taxi
- Driverless Robo Bus
- Human Driver vehicles
L2+ Products: Semi-Autonomous Highway/Urban Features
- Human Driver vehicles
Self-Driving Two Separate Branches…
Comfort Driving with safety guarantees
Where are we heading…
• Key differentiation through XPENG’s intelligent EV strategy
• Focus on in-house software solutions running on leading
computer hardware platforms
• Focus on enhancing driver assistance features towards L3
capabilities in 2020
• Focus on feature optimization for Chinese market and
Chinese consumers
8
9
Roadmap for XPENG Motors’ Autonomous Driving Solution XPilot
2018-2019 2020 2021-2022 2024-2025
CarModel
Autonomous
Driving
Achievement
&Goal
• Xpilot 2.0:
Full scenario auto
parking + LCC + ACC
• Xpilot 3.0:
Super Memory Parking
+ Highway Autopilot
• Xpilot 4.0:
AVP + Highway/Urban
Autopilot
• Xpilot x.0:
Full Automation in
limited scenarios
(highway, parking, etc.)
• First generation
solution for
commercialization
with partial self
developed IPs
• Fully self developed
end-to-end SW IP
• Next gen HW
platform with
optimized sensor cost
• First level-4 solution
in consumer vehicle
A Class SUV (G3) B Class Sedan (E28) B Class SUV
10
XPENG G3
By AI,
irregular lots
By memory
Roadmap for XPENG Motors’ Autonomous Driving Solution-XPilot
parking functions
XPENG E28
Alleviate people’s parking anxiety in parallel, vertical, perpendicular, irregular lots…
parking functions
By voice
By Vision
By Ultrasonic
11
Use Case Example: Super Memory Parking
12
Use Case Example: Highway Auto-Pilot
13
2km ahead
Xpilot’s Surrounding Reality
Ego
Traffic
Point of interest
Road side object
Lane condition
14
Agenda
1. XPENG Motors introduction and background
2. Roadmap for XPENG Motors’ Autonomous Driving Solution
3. System architecture design powered by Xavier
15
Driving
Big Data
AI and
Perception
High-perf
chips
Customized
Sensor
Structure
- -
Chinese Traffic Patterns
Local adaption is a must
- -
• Chinese style lane change and close cutting in
• Very large variety of obstacles
• Dense traffic
• Irregular Pattern
16
Challenge: Perception & Creating a Representative 3D Environmental Model
Compute (MP/s)
Quality raw sensor input with cost
constraints
2
1
Preponderance of training data
(especially corner cases)
3
17
Challenge: Perception & Creating a Representative 3D Environmental Model
Real stream taken from E28 camera sets; ISP developed in house;
18
Challenge: Extending to Day/Night, All Weather, All locations
1) Time (Location of Sun: Day/Night/NotDawn/NotDusk)
2) Select Weather Conditions
3) Availability of GNSS & RTK
4) Place (GeoFenced)
5) Select Speed Range
6) Road/Traffic Conditions
Sensing/Perception challenges (due
to insufficient quality of raw data
and/or compute power)
Mapping challenges
Driving policy
challenges
Ultrasonic Sensors
(6 in Front)
Tri-focal Cameras
(2M pixels, FOV 28,
52, 100, 60/15fps)
Surround-view
Camera (1 at Back)
Rear-view Camera
(2M pixels,FOV 52,
30fps)
In-Car Intelligent
Camera
Radars
(5 in Total)
Surround-view Camera
(1 in Front)
Ultrasonic Sensors
(6 at Back)
Surround-view
Camera (2 on Sides)
Cut-In Prevention
Camera
1M pixels, 60fps)Rear-view Camera
(1M pixels, 30fps)
Sensor System for E28 (Mass production in 2020)
19
Note: We use first vehicle G3’s picture as E28’s design is not revealed yet.
20
XPilot Camera Coverage and Functions for E28
360 coverage; Designed to cover traffic scenarios in China
HFOV 100
HFOV 52
21
System Differentiators
22
1828x948@15fps
(for Highway Driving)
457x237@60fps
(for Urban Driving)
Long Range CNN
(Perception at > 200m)
Close Proximity CNN
(Responds x4 faster)
Cameras
Variable fps Network for China Specific Drive Conditions
(Close Vehicle Proximity and Abrupt Cut-ins)
1
23
HighwayUrban Canyon
START POINT /
END POINT
55 Km route (Approx. 2.5hr)
2.5hrs x 3600 x 1Hz = 9,000 samples
GuangzhouRoute8
Absolute
Localization
Reference
Antenna +
GNSS/INS
2 GNSS / INS
20cm absolute-localization-accuracy at 95% (2 σ)
20cm absolute-localization-accuracy for at 95% (2 σ) of Highway route in white
24
Front Radars
Right Front
Corner Radar
Left Front
Corner Radar
Left Rear
Corner Radar
Right Rear
Corner Radar
3 RADAR
Five Radars of High Angular Resolution: 2x improvement
5th Generation of Radar Technology
25
Industry Leading Computing Platform
XPilot Unit (XPU) powered by NVIDIA Xavier
NVIDIA Xavier
Safety MCU
Camera Des
Camera Des
Camera Des
Memory
EMMC
Storage
CAN
Transceiver
Ethernet
Switch
Power
Management
CAN Ethernet
4
26
Why NVIDIA Xavier?
0
5
10
15
20
25
30
35
Computing Power (TOPS)
Deep Learning Computing Throughput
NVIDIA Xavier FPGA NVIDIA Parker TI TDA2xx
TOPS
27
XPU Software Stack
NVIDIA Xavier SoC Aurix MCU
Hypervisor
NVIDIA
Foundations
QNX OS and BSP/Drivers
NVIDIA System Software NvMeida/CUDA/CuDNN/TensorRT …
NVIDIA Drive Works
XPilot Frameworks
(Project Eagle)
NVIDIA
Tier-1
AUTOSAR
OTA Data
Camera
Service
CAN
Service
(Radar,
IMU/GNSS,
Vehicle IO)
XPilot
Autonomous
Driving
Applications
Perception,
Localization
Prediction, Path
Planning
Control, …
XPU HW NVIDIA/Tier- SW XPU Platform SW XPilot SW
Visualization Launcher
Monitor
Diagnostic
Service
Xpilot
Safety
App
28
Driving hierarchy w/ safety intervention
AEB
Driver: Human Driver: XPilot
Collision
Warnings
5 SAFETY
Independent AEB System
29
Summary
• XPENG Motors develop FUN, CONNECTED and INTELLIGENT EVs tailored for the
young and tech-savvy Chinese millennials
• Roadmap for XPENG Motors’ Autonomous Driving Solution
• System architecture design powered by Xavier
30
E28 is coming…
To be revealed at Auto Shanghai April 16

Xpeng Motors' P7's self-driving roadmap and system design

  • 1.
    0 Developing Autonomous DrivingEVs for the China Market XPENG Motors’ Approach Junli Gu VP of Autonomous Driving, XPENG motors March 18, 2019
  • 2.
    1 Agenda 1. XPENG MotorsIntroduction and Background 2. Roadmap for XPENG Motors’ Autonomous Driving Solution 3. System Architecture Design Powered by Xavier
  • 3.
    2 Who is XPENG? Wedevelop FUN, CONNECTED and INTELLIGENT EVs tailored for the young and tech-savvy Chinese Millennials 2 • Aimed at young Internet users in China, with autonomous driving and intelligent connection as the core differences • 2014: Founded in Guangzhou, China • Dec 2018: Mass production and official launch of G3 (electrical SUV) • Q1 2019: Mass production of E28 (electrical sedan)
  • 4.
    3 Strong R&D capabilitiesand investments Shanghai Intelligent Connected Vehicle Navigation R&D Silicon Valley & San Diego Machine Learning Autonomous Driving R&D Guangzhou Vehicle and VCU R&D V2X Testing Trial and Production Lead-in Beijing Artificial Intelligence R&D R&D footprint Our R&D headcount accounts for over 70% of our total number of employees.
  • 5.
    4 Your Smart DrivingAssistant Connectivity Xmart OS X-pilot Artificial Intelligence
  • 6.
  • 7.
    6 Agenda 1. XPENG MotorsIntroduction and Background 2. Roadmap for XPENG Motors’ Autonomous Driving Solution 3. System Architecture Design Powered by Xavier
  • 8.
    7 L3 Products: Highly-AutonomousHighway Features L4 Products: Fully-Autonomous Urban & Highway Features - Driverless Robo Taxi - Driverless Robo Bus - Human Driver vehicles L2+ Products: Semi-Autonomous Highway/Urban Features - Human Driver vehicles Self-Driving Two Separate Branches…
  • 9.
    Comfort Driving withsafety guarantees Where are we heading… • Key differentiation through XPENG’s intelligent EV strategy • Focus on in-house software solutions running on leading computer hardware platforms • Focus on enhancing driver assistance features towards L3 capabilities in 2020 • Focus on feature optimization for Chinese market and Chinese consumers 8
  • 10.
    9 Roadmap for XPENGMotors’ Autonomous Driving Solution XPilot 2018-2019 2020 2021-2022 2024-2025 CarModel Autonomous Driving Achievement &Goal • Xpilot 2.0: Full scenario auto parking + LCC + ACC • Xpilot 3.0: Super Memory Parking + Highway Autopilot • Xpilot 4.0: AVP + Highway/Urban Autopilot • Xpilot x.0: Full Automation in limited scenarios (highway, parking, etc.) • First generation solution for commercialization with partial self developed IPs • Fully self developed end-to-end SW IP • Next gen HW platform with optimized sensor cost • First level-4 solution in consumer vehicle A Class SUV (G3) B Class Sedan (E28) B Class SUV
  • 11.
    10 XPENG G3 By AI, irregularlots By memory Roadmap for XPENG Motors’ Autonomous Driving Solution-XPilot parking functions XPENG E28 Alleviate people’s parking anxiety in parallel, vertical, perpendicular, irregular lots… parking functions By voice By Vision By Ultrasonic
  • 12.
    11 Use Case Example:Super Memory Parking
  • 13.
    12 Use Case Example:Highway Auto-Pilot
  • 14.
    13 2km ahead Xpilot’s SurroundingReality Ego Traffic Point of interest Road side object Lane condition
  • 15.
    14 Agenda 1. XPENG Motorsintroduction and background 2. Roadmap for XPENG Motors’ Autonomous Driving Solution 3. System architecture design powered by Xavier
  • 16.
    15 Driving Big Data AI and Perception High-perf chips Customized Sensor Structure -- Chinese Traffic Patterns Local adaption is a must - - • Chinese style lane change and close cutting in • Very large variety of obstacles • Dense traffic • Irregular Pattern
  • 17.
    16 Challenge: Perception &Creating a Representative 3D Environmental Model Compute (MP/s) Quality raw sensor input with cost constraints 2 1 Preponderance of training data (especially corner cases) 3
  • 18.
    17 Challenge: Perception &Creating a Representative 3D Environmental Model Real stream taken from E28 camera sets; ISP developed in house;
  • 19.
    18 Challenge: Extending toDay/Night, All Weather, All locations 1) Time (Location of Sun: Day/Night/NotDawn/NotDusk) 2) Select Weather Conditions 3) Availability of GNSS & RTK 4) Place (GeoFenced) 5) Select Speed Range 6) Road/Traffic Conditions Sensing/Perception challenges (due to insufficient quality of raw data and/or compute power) Mapping challenges Driving policy challenges
  • 20.
    Ultrasonic Sensors (6 inFront) Tri-focal Cameras (2M pixels, FOV 28, 52, 100, 60/15fps) Surround-view Camera (1 at Back) Rear-view Camera (2M pixels,FOV 52, 30fps) In-Car Intelligent Camera Radars (5 in Total) Surround-view Camera (1 in Front) Ultrasonic Sensors (6 at Back) Surround-view Camera (2 on Sides) Cut-In Prevention Camera 1M pixels, 60fps)Rear-view Camera (1M pixels, 30fps) Sensor System for E28 (Mass production in 2020) 19 Note: We use first vehicle G3’s picture as E28’s design is not revealed yet.
  • 21.
    20 XPilot Camera Coverageand Functions for E28 360 coverage; Designed to cover traffic scenarios in China HFOV 100 HFOV 52
  • 22.
  • 23.
    22 1828x948@15fps (for Highway Driving) 457x237@60fps (forUrban Driving) Long Range CNN (Perception at > 200m) Close Proximity CNN (Responds x4 faster) Cameras Variable fps Network for China Specific Drive Conditions (Close Vehicle Proximity and Abrupt Cut-ins) 1
  • 24.
    23 HighwayUrban Canyon START POINT/ END POINT 55 Km route (Approx. 2.5hr) 2.5hrs x 3600 x 1Hz = 9,000 samples GuangzhouRoute8 Absolute Localization Reference Antenna + GNSS/INS 2 GNSS / INS 20cm absolute-localization-accuracy at 95% (2 σ) 20cm absolute-localization-accuracy for at 95% (2 σ) of Highway route in white
  • 25.
    24 Front Radars Right Front CornerRadar Left Front Corner Radar Left Rear Corner Radar Right Rear Corner Radar 3 RADAR Five Radars of High Angular Resolution: 2x improvement 5th Generation of Radar Technology
  • 26.
    25 Industry Leading ComputingPlatform XPilot Unit (XPU) powered by NVIDIA Xavier NVIDIA Xavier Safety MCU Camera Des Camera Des Camera Des Memory EMMC Storage CAN Transceiver Ethernet Switch Power Management CAN Ethernet 4
  • 27.
    26 Why NVIDIA Xavier? 0 5 10 15 20 25 30 35 ComputingPower (TOPS) Deep Learning Computing Throughput NVIDIA Xavier FPGA NVIDIA Parker TI TDA2xx TOPS
  • 28.
    27 XPU Software Stack NVIDIAXavier SoC Aurix MCU Hypervisor NVIDIA Foundations QNX OS and BSP/Drivers NVIDIA System Software NvMeida/CUDA/CuDNN/TensorRT … NVIDIA Drive Works XPilot Frameworks (Project Eagle) NVIDIA Tier-1 AUTOSAR OTA Data Camera Service CAN Service (Radar, IMU/GNSS, Vehicle IO) XPilot Autonomous Driving Applications Perception, Localization Prediction, Path Planning Control, … XPU HW NVIDIA/Tier- SW XPU Platform SW XPilot SW Visualization Launcher Monitor Diagnostic Service Xpilot Safety App
  • 29.
    28 Driving hierarchy w/safety intervention AEB Driver: Human Driver: XPilot Collision Warnings 5 SAFETY Independent AEB System
  • 30.
    29 Summary • XPENG Motorsdevelop FUN, CONNECTED and INTELLIGENT EVs tailored for the young and tech-savvy Chinese millennials • Roadmap for XPENG Motors’ Autonomous Driving Solution • System architecture design powered by Xavier
  • 31.
    30 E28 is coming… Tobe revealed at Auto Shanghai April 16