Applying AI to Mapping

Transcription

1 Applying AI to Mapping Dr. Ryan Wolcott Manager, Simultaneous Localization and Mapping (SLAM) 1

2 TRI: Who We Are Our mission is to improve the quality of human life through advances in artificial intelligence, automated driving, and robotics. We are dedicated to creating a world where everyone, regardless of age or ability, can work in harmony with technology to enjoy a better life. Through innovations in AI, we will develop a car incapable of causing a crash, regardless of the actions of the driver. We will also develop technology for vehicles and robots to help people enjoy new levels of independence and mobility in their daily lives. 2

3 SAE Levels of Automation J3016 Level 1 Levels 4/5 One degree of freedom is controlled, e.g., longitudinal Adaptive Cruise Control Human is always a passenger; the only distinction between 4 and 5 is operational domain (i.e., geofenced vs. unrestricted) NO A U T O M AT I O N D R I V E R A S S I S TA N C E PA R T I A L A U T O M AT I O N C O N D I T I O N A L A U T O M AT I O N H I G H A U T O M AT I O N F U L L A U T O M AT I O N Level 2 Level 3 Readiness for hand-off requires constant vigilance from driver that may not be sustainable over time Difficult for vehicle to ensure driver has sufficient warning to re-engage in time for hand-off 3

4 Automated Driving Approach: One System, Two Modes The Ultimate TSS L4/L5 Equivalent Driver always engaged, but vehicle monitors and intervenes to help prevent collisions Fully autonomous driving system engaged at all times Builds on similar hardware and software development as fully-autonomous Chauffeur Staged commercial release, likely beginning with shared mobility fleets 4

5 Core Elements of Automated Vehicles PERCEPTION Interprets what the vehicle sees PREDICTION Anticipates what others might do PLANNING Decides an action the vehicle will take 5

6 Core Elements of Automated Vehicles PERCEPTION PREDICTION AI PLANNING 6

7 Core Elements of Automated Vehicles Radar PERCEPTION PREDICTION Cameras LIDAR AI PLANNING 7

8 The Use of Your navigation system operates from onboard maps PROBLEM: Roads are always under construction Most maps are outdated the moment they re created Automated vehicles use more sophisticated maps plus localization to simplify perception LIDAR BUT they have similar map issues Perception Radar Cameras 8

9 The Use of LIDAR Perception Radar Cameras 9

10 The Use of Lane center/adjacency Stop lines Intersection connectivity LIDAR Perception Radar Cameras 10

11 The Use of Lane center/adjacency Stop lines Intersection connectivity Traffic lights Crosswalks LIDAR Perception Radar Cameras 11

12 The Use of Lane center/adjacency Stop lines Intersection connectivity Traffic lights Crosswalks Offline trajectory optimization Etc. LIDAR Perception Radar Cameras 12

13 Common Approach to Level 4/5 What do we do with HD maps? HD Localization Lane-level Map for Navigation Planning Stack 13

14 Challenges of Level 4/5 Many of these challenges occur when HD map is inconsistent with live perception! Changes to road surface markings Police redirect traffic Map errors 14

15 TRI HD Mindset HD Localization Lane-level Map for Navigation Planning Stack 15

16 TRI HD Mindset HD Live Perception TRI Semantic Segmentation Localization AI Lane-level Map Lane-level Map for Navigation for Navigation Planning Stack TRI Feature Extraction 16

17 TRI Map Mindset Guardian Connection Chauffeur Capability SD HD 17

18 TRI Map Mindset Guardian Connection 1) Some functionality everywhere, spectrum of map inputs Chauffeur 2) Learn map over time Capability Guardian HD Map Coverage Guardian SD Map Coverage SD HD Time 18

19 TRI s Automated Vehicle Testing Platform Platform 3.0 One of the most perceptive test cars on the road 200m range 360-degree coverage New integrated design that s sleek and elegant Weatherproof rooftop panel seamlessly conceals equipment 19

20 20

21 Closing Perspective AI is the brain of the automated vehicle system Role of the map Challenge for Level 4/5 automation: inconsistency between map and live perception Possible solution: fuse AI with perception and maps Unique approach enables TRI s exploration of Guardian mode 2017 Toyota Research Institute. Proprietary and confidential. Do not distribute.

22 Thank You 22