Turing株式会社のEN_2015_Machine Learning Engineer (Autonomous Driving — World Model / Video Generation Model Development)の求人

-About the role-
◆This role is open to engineers with a machine learning background who have worked on large-scale development of video generation models or foundation models, as well as the MLOps / data infrastructure, distillation, optimization, and on-device deployment that supports them. (Experience in autonomous driving / CV / robotics is a plus.)◆

At Turing, we are developing an End-to-End autonomous driving model that takes input from vehicle-mounted cameras and directly controls the vehicle.
As we work toward full self-driving, we are in an exploratory phase
— combining imitation learning with the rapid advances in video generation and world model research to actively find what works.

In particular, large-scale pretraining on not only driving data but also general video data (video generation, self-supervised learning, etc.) has been shown to significantly boost downstream task performance (behavior prediction, planning, etc.), and this trend is growing stronger.
With this in mind, Turing is pursuing a World Action Model (WAM) approach — a unified pipeline that spans modern model families such as video generation models, image/video foundation models (e.g., DINOv3, V-JEPA-style concepts), and world models — from training at scale, through validation and downstream task integration, to distillation, quantization, inference optimization, and final deployment on real vehicles.

We are looking for members who can take these ambitious research directions all the way to something that actually works in the real world.

■What you will work on
・ML development centered on WAM (World Action Model) for autonomous driving
・Large-scale pretraining (scaling training with driving data + general video data, etc.)
・Modeling, implementation, and validation using video generation models and world models
・Validation and application of image/video foundation models based on self-supervised learning (e.g., DINOv3, V-JEPA-style, etc.)
・Application to downstream tasks (e.g., behavior prediction, planning), evaluation design, and improvement
・Model compression and acceleration via distillation, quantization, and inference optimization
・Building and improving experimental infrastructure (data pipelines, reproducibility, experiment management, model operations)
・Literature review and implementation validation in related areas (Transformers, robotics, world models, etc.)

■Enjoy being at the frontier of Physical AI
Giving AI a physical presence and enabling it to deliver value in the real world — autonomous driving is exactly where humanity is pushing this frontier today. You will need to build unique ML pipelines while leveraging the knowledge already accumulated within the company. We are looking for someone who can drive development in a domain with almost no existing reference points.

■Test your model in the real world
Our development cycle: Build dataset & model → Drive test → Analyze experiment logs → Manage model. You will iterate on your models by experiencing them firsthand in a real vehicle. Use feedback from the physical world to drive your development forward.

■Who is thriving in this role
・Engineers with strengths in robotics, world models, or autonomous driving (behavior prediction, planning, etc.) who have led model development
・Engineers who have pursued large-scale data preprocessing, filtering, and data quality design, and have achieved training reproducibility and scaling in practice
・Engineers from research labs or corporate research teams who have taken exploratory topics all the way from implementation → validation → improvement to tangible results
・Engineers who can quickly catch up with the work of leading researchers and recent papers, reproduce and extend them, and connect the results to product or on-vehicle validation

■Who we are looking for
・Driven to build a world-class company
・Self-starter who takes initiative on everything
・Humble, with genuine empathy for others
・Flexible and excited by rapid organizational and business growth
・Growth-oriented mindset
・Resilient — able to find joy even in tough challenges
◆Tech stack
・Language：Python
・Libraries：PyTorch、Transformers、Lightning、Accelerate、DeepSpeed/FSDP、FlexAttention
・Middleware：Slurm
・Dev environment：Large-scale GPU clusters built on AWS and GCP
・Platform：Jetson、Linux

【LEARN MORE】
▼Company website
https://tur.ing/

▼Turing Tech Blog
https://zenn.dev/p/turing_motors

▼Turipo（Owned media / Web Magazine）
https://tur.ing/turipo

▼Turing TechTalk #7 E2E Autonomous Driving Development Process
https://youtu.be/KHqGVkIhYp4?feature=shared

【Application notes】
· Please submit your resume and work history in PDF format.
· Do not include salary information (current or desired) in your application documents or entry form.
· Our HR team will discuss compensation separately during the selection process.

■Required qualifications
・Practical experience in model development using machine learning / deep learning
・Experience implementing and operating training code using PyTorch or similar
・Understanding of Transformer-based architectures, with experience implementing or modifying them
・Foundational understanding of large-scale training (distributed training, optimization, training stabilization, experiment management), with hands-on experience in at least one area
・Ability to read research papers and technical documents, and reproduce / validate their findings
・Japanese language proficiency (JLPT N2 or equivalent)

■Preferred qualifications
▼Generative & foundation models
・Exploration of video generation model training recipes (Diffusion / Flow Matching, long-sequence training, self-forcing-style recovery training)
・Research and development experience with image/video tokenizers for generative modeling downstream tasks
・Experience with pretraining or transfer learning of image/video foundation models via self-supervised learning (e.g., DINOv3, V-JEPA, etc.)
・Large-scale pretraining for robot tasks using Latent Action Models
・Research and development related to visual geometry foundation models such as Feed-Forward 3DGS or DepthAnything3
▼Distillation & deployment
・Experience with model compression techniques including distillation, quantization, and inference optimization
・Experience fine-tuning and distilling video generation models for robot tasks and deploying them to real hardware
・Experience meeting inference requirements (latency / throughput / memory) under constraints such as in-vehicle or edge environments