Automated Driving Advanced Development Intern, Machine Learning Research

The Team

The Automated Driving Advanced Development division bridges TRI research and Toyota products. In partnership with Woven by Toyota, the team is researching and developing an end-to-end learned driving stack that maps sensor inputs directly to planned trajectories.

The Internship

We’re seeking a research intern to investigate how driving policies post-trained in fast, low-fidelity simulation can transfer to higher-fidelity simulation and real-world driving. This is an open research problem: differences in sensor inputs, vehicle dynamics, and other agents can limit transfer, and policies may learn simulator-specific behavior or forget previously learned skills.

You’ll lead an independent research project exploring approaches to sim-to-sim and sim-to-real transfer for sensor-based driving models. You’ll work with Toyota driving data, generative base models, and closed-loop simulation infrastructure, with the goal of producing work suitable for a conference or workshop submission.

The Opportunity

As a research intern on AD2, you will own a standalone research project in sim-to-sim and sim-to-real domain mapping, with the opportunity to:

• Drive your own research: Shape the approach, run experiments, and analyze results, with the independence to set your own direction.

• Publish your findings: The project is designed to produce results suitable for a conference or workshop publication.

• Shape a real autonomy model: The tools and insights you create will help the team post-train our autonomy model more effectively, with the potential to influence the autonomy stack eventually deployed in Toyota vehicles.

• Collaborate with experienced researchers: Work closely with a small team of industry-leading researchers with deep experience in robotics and autonomous driving.

• Work at scale: Use the same large-scale Toyota driving data, base models, and simulation infrastructure the team uses every day.

You can learn more about our work on the Automated Driving Advanced Development team page.

• Conduct research to advance the state of the art in sim-to-sim and sim-to-real domain mapping, enabling driving policies post-trained in fast, low-fidelity simulation to transfer to higher-fidelity simulation and real-world driving.
• Design and run experiments using large-scale Toyota driving data, the team’s generative base models, and closed-loop simulation infrastructure.
• Perform closed-loop evaluations across simulators of varying fidelity, and develop metrics that measure how well post-trained behaviors carry over between domains.
• Collaborate with researchers and engineers to integrate the tools and findings from your project into the team’s model development and post-training workflows.
• Present progress and results regularly to the team, and incorporate feedback into the research direction.
• Write up your findings as a paper for submission to a conference or workshop by the end of the internship.

• Currently pursuing a Ph.D. or master’s degree in Computer Science, Robotics, Machine Learning, Artificial Intelligence, Engineering, or a related field.
• Proficiency in Python and ML frameworks such as PyTorch for implementing and evaluating research ideas, along with an understanding of version control, testing, and software engineering fundamentals.
• Research experience in one or more of the following: sim-to-real transfer, domain adaptation, reinforcement learning, imitation learning, generative models such as diffusion or flow matching, or autonomous driving.
• Demonstrated research ability, such as publications, preprints, or substantial independent research projects.
• Ability to work independently on an open-ended research problem, and to communicate methods and results clearly.
• Able to work in a hybrid in-office arrangement at TRI’s headquarters in Los Altos, California, for the full 12-week internship.

• Experience with driving or robotics simulators such as CARLA or NVIDIA DriveSim, or with sensor simulation and neural rendering.
• Experience with world models or generative models for video or scene reconstruction.
• Experience post-training policies with reinforcement learning, DAgger, or preference-based methods.
• Hands-on experience working with real-world sensor data, such as video, lidar, IMU, or radar.
• Experience with large-scale distributed training and experiment tracking tools such as Weights & Biases.
• Publications at venues such as CoRL, ICRA, RSS, NeurIPS, ICML, or CVPR.