An Edge-Cloud System Model for Autonomous Vehicles Yu Sasaki - PowerPoint PPT Presentation

An Edge-Cloud System Model for Autonomous Vehicles Yu Sasaki † ,Tomoya Sato † , Hiroyuki Chishiro † ,Tasuku Ishigooka § , Satoshi Otsuka § , Kentaro Yoshimura § , Shinpei Kato †‡ † Dept. Computer Science, The University of Tokyo, Japan § Hitachi, Ltd., Japan ‡ Tier IV, Inc., Japan 2019/11/4

Background In Autonomous Driving... • High Computational Tasks Scan Data • Self-localization Map Data • Path Planning • Multiple Sensing Inputs • Location ：GPS / IMU • Scans ：Cameras / LiDAR Self-localization • Map Data：Point-cloud Map / ADAS Map Path Planning 1 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Automotive Embedded Systems Advance in embedded hardware Embedded Hardware Commercial PC Pros Cons • Low power consumption • Insufficient computation power • Small • Low Upgradability 2 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Distributed Computing Approach Vehicular Ad Hoc Networks (VANET) [Benslimane+,2004] • Creating ad-hoc network within vehicles • Communicating road information • Distributed computing • Information pass-thru • Cons • Not Ideal with a few vehicles • Interference from non-VANET vehicles (Obstacle Shadowing) [Meireles+,2010] A. Benslimane et al : “Optimized Dissemination of Alarm Messages in Vehicular Ad-Hoc Networks (VANET)” , 2004 R. Meireles, et al : “Experimental Study on the Impact of Vehicular Obstructions in VANETs” , 2010 3 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Edge-Cloud Computing Approach In other fields... 4 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Why few edge-cloud model? Why now? • Requires higher network bandwidth • Advance in next-gen wireless network It has become more practical to consider edge-cloud model 5 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Contribution • Edge-cloud system model for autonomous vehicles • Effectiveness of presented model • Round-trip execution time • Improvement rate • Case study using Autoware [Kato+,2015] S Kato et al : “An Open Approach to Autonomous Vehicles”, 2015 6 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Round-trip Execution Time • Execution time on edge machine • Execution time on cloud server • Communication cost between edge and cloud • Communication delay due to packet loss 7 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

System Model Model Notation • Task : 𝑈 = {𝑘 % , … , 𝑘 ( } Num of jobs : 𝑂 • Jobs run on cloud : 𝐾 , Jobs run on edge : 𝐾 - • Job pair with edge-cloud communication : 𝐹 = {(𝑞, 𝑟), … } • Execution time of 𝑘 3 on cloud : 𝑓 ,,5 6 Execution time of on 𝑘 3 edge : 𝑓 -,5 6 • Bandwidth : 𝐶 Packet loss : 𝑔 • Traffic : 𝑢𝑠𝑏𝑜𝑡 >,? average transmission speed : 𝑐 >,? 𝑢𝑠𝑏𝑜𝑡 >,? 𝑇 = ∑ 𝑓 ,,5 6 + ∑ 𝑓 -,5 6 + (1 + 𝑔) ∑ ∑ 𝑐 I,5 ≤ 𝐶 𝑐 >,? 5 6 ∈D E 5 6 ∈D G (>,?)∈- (>,?)∈- 𝑢𝑠𝑏𝑜𝑡 >,? 𝑇 = ∑ 𝑓 ,,5 6 + ∑ 𝑓 -,5 6 + (1 + 𝑔) ∑ ∑ 𝑐 I,5 > 𝐶 𝐶 5 6 ∈D E 5 6 ∈D G (>,?)∈- (>,?)∈- 8 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Experiment • Experimented with wired Ethernet connection (simulating 5G / Wireless 6 speed) • Constructed model assuming no packet loss 9 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Experiment • Evaluated using simulation mode on Autoware [Kato+,2015] • ROS-based open-source software (https://www.autoware.ai/) • DRIVE PX2 as edge machine, desktop PC as cloud server • Evaluation • Self-localization • Round-trip execution time • Deadline miss rate • Path planner • Execution time • Each evaluation on • Edge (standalone) • Cloud(standalone) • Edge-cloud S Kato et al : “An Open Approach to Autonomous Vehicles”, 2015 10 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

DRIVE PX2 • 2 SoC board • Denver(x2)= Cortex-A57(x4) • Pascal GPU • 512 CUDA core • 4GB Mem • 6GB Memory • 64GB storage • 80W https://devblogs.nvidia.com/wp-content/uploads/2018/04/self-driving-car-drive-px-autochauffeur-516-ud-1.png 11 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Self-localization ② ① ② /filtered_points ① /points_raw 12 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Grid Map Filter ③ ④ ① ⑤ ① /points_raw ③ /points_no_ground ④ /realtimecostmap ⑤ /distance_transform 13 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Path Planning ⑤ ⑥ /current_pose 14 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Experiment Environment Machine Environment Edge(DRIVE PX2) Cloud CPU Cortex-A57(x4), Denver(x2) Core i7-8700 CPU Freq. 2.0GHz(A57), 2.0GHz(Denver) 3.2GHz Memory 6GB 32GB GPU Memory 4GB 8GB(GTX1080) CUDA Core 512 1280 Linux kernel 4.9.38 4.16.5 ROS kinetic Simulation Dataset • 200sec drive data acquired in Nagoya University • Deadline • Self-localization : 100ms • Path-planning : 2,000ms 15 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Simulation Data 16 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Results • Measured execution time of self-localization for 2,000cycles (200sec) • Average: (Edge) 55.4ms-> (Edge-Cloud) 35.1ms • Improved 57% • WCET: (Edge) 348ms -> (Edge-Cloud) 103ms • Improved 3.5x 17 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Results • Measured 100 sets (each set consists of 2,000 cycles) • Total of 200,000 cycles • Worst deadline rates (Edge) 3% ->(Edge-Cloud) 0.20% • Only 3 misses out of 200,000 cycles 18 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Results • Measured execution time of path planner for 100 cycles • In some cases 10x faster than standalone edge machine • Drastic difference in goals that requires curving path 19 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Result 20 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Results • Measured execution time of path planner for 100 cycles • In some cases 10x faster than edge machine • Drastic difference in goals that requires curving path 21 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.

Conclusion • We presented edge-cloud computing model for autonomous driving • Even with the communication cost, our model is faster and more stable • Future Work Tests in real world environment (5G / Wireless 6) Questions? E-Mail : Yu Sasaki (yu.sasaki@pf.is.s.u-tokyo.ac.jp) 22 An Edge-Cloud System Model for Autonomous Vehicles (PPNIV '19) Yu Sasaki et al.