Eiji Nakatsugawa

Joined Woven Core as a seconded of TOYOTA to learn deeply about the "Software First" development. In charge of releasing the autonomous driving object detection's new feature for software update, and involved in coding/collecting data for the production-level recognition functions. Also leading the project of utilizing the "20TM/TSS3 Market Data Collection" for the next generation recognition system development. The achievements through this project are described below:

• Implemented transform chain feature to improve occlusion/truncation classification performance
• Collected warning sign dataset by using detection network
• Estimated vehicle speed threshold from market data collection to optimize the trigger specification
• Implemented Super-Resolution pipeline to improve low-resolution camera images coming from the market data collection
• Designed new trigger to efficiently search specific corner cases from market data collection
• Prepared "Offline DNN inference based tagging pipeline" to have efficient platform to collect data easily collaborating with MDC team in Woven Alpha.
• Prepared ML-Pipeline using Kubeflow Pipeline to automate the production-level development processes to reduce the man-hour collaborating with AreneAI team in Woven Alpha.
• 1 patent application

In charge of improving the efficiency of vehicle development using machine learning, and completed planning proposals and implementation of methods that combined domain knowledge (conventional technology and know-how) and machine learning. In addition, we improved the environment and planned a digital human resource development program including machine learning. The achievements through this project are described below:

• Build hydraulic engine mount component model with non-linear characteristics by combining machine learning methods (PINNs: Physics Informed Neural Network and RNN: Recurrent Neural Network)
• Prepared education program and environment for growing AI/ML engineer called "Digital Dojo"
• Estimate intermediate characteristics of vehicle system by statistical causal search using Bayesian network and LiNGAM: Linear Non Gaussian Model
• Estimate sensory evaluation KPI of drivability performances from time series data of vehicle's multi-points acceleration
• 3 patent application
• Published paper in 2020 JSAE (Society of Automotive Engineers of Japan, Inc.)
• Published paper in 2021 IPSJ (Information Processing Society of Japan, Inc.)
• Published paper in 2021 JSME iDECON (https://www.jsme.or.jp/conference/iDECON-MS2021/doc/program.html)
• Complete NABLAS iLect DL4E (AI/ML education program)
• Kaggle 2020 lish-moa: silver medal 121th/4384

In charge of advanced development of vehicle vibration and noise performance, I have been involved in implementation and promotion as a leader in the development process KAIZEN of actuator operating noise involving suppliers (DENSO, ADVICS). Among them, MBD: Model-Based Development predicted the performance in the simulation world at the initial stage of planning, and proposed better design specifications, which led to the reduction of man-hours and costs. The achievements through this project are described below:

• Build "Brake Actuator System" simulation model including software and hardware
• Build the development process architecture including suppliers (DENSO, ADVICS, TMC)
• 3 patent application
• Reduce 60% of the man-hour/cost comparing with previous development process
• Published paper in 2019 JSME (The Japan Society of Mechanical Engineers; https://www.jsme.or.jp/20210305-2/)
• Published paper in 2019 JSAE (Society of Automotive Engineers of Japan, Inc.)