Research Engineer, Computer Vision - Learning From Videos (LFV)

The Mission

Make general-purpose robots a reality.

The Challenge

We envision a future where robots assist with household chores and cooking, aid the elderly in maintaining their independence, and enable people to spend more time on the activities they enjoy most. To achieve this, robots need to operate reliably in messy, unstructured environments. Our mission is to answer the question: what will it take to create truly general-purpose robots that can accomplish a wide variety of tasks in settings like human homes, with minimal human supervision? We believe that the answer lies in cultivating large-scale datasets of physical interaction from a variety of sources and building on the latest advances in machine learning to learn general-purpose robot behaviors from this data.

The Team

The Learning From Videos (LFV) team in the Robotics division develops foundation models that leverage large-scale multi-modal data (RGB, depth, flow, semantics, actions, tactile, audio, etc.) from multiple domains (driving, robotics, indoors, outdoors, etc.) to power downstream embodied AI tasks. Our topics of interest include Video Generation, World Models, 4D Reconstruction, Multi-Modal Models, Multi-View Geometry, Data Augmentation, and Video-Language-Action models, with a primary focus on embodied applications. We are making progress on some of the hardest scientific challenges around spatio-temporal reasoning, and how it can lead to the deployment of autonomous agents in real-world unstructured environments, across both robotics and driving domains.

The Opportunity

Our team is looking for a Research Engineer to own and drive the core data and model infrastructure that powers our research. As our foundation models scale in both data diversity and model complexity, we need a strong engineer who can bridge the gap between research ideas and production-grade systems. This is not a traditional software engineering role; you will work directly alongside research scientists, understand the research deeply enough to make independent technical decisions, and play a key role in enabling the team to move faster and train better models.

As a Research Engineer, you will be responsible for building and maintaining the infrastructure that ingests, unifies, and serves heterogeneous multi-modal datasets at scale; supporting and optimizing large-scale distributed training of diffusion and transformer models; and developing tools and pipelines that accelerate the research-to-results cycle. You will work closely with researchers to prototype new ideas, run experiments, and help ship our most successful models toward real-world applications.

Build and maintain scalable pipelines for ingesting, converting, validating, and serving heterogeneous robotics and vision datasets (multi-view, multi-modal, multi-embodiment, etc.) into unified training-ready formats. Track and integrate new public and internal datasets as they become available.

Support and optimize large-scale distributed training of foundation models (diffusion transformers, video generation models) on multi-GPU and multi-node clusters. Manage experiment workflows, profiling, debugging, and hyperparameter sweeps.

Collaborate directly with research scientists to implement, iterate on, and evaluate new architectures, objectives, and training strategies. Translate research prototypes into clean, maintainable, reusable code.

Develop tools for dataset inspection, experiment tracking, model evaluation, GPU resource management, and visualization. Automate repetitive workflows to improve team velocity.

Work with other TRI teams and Toyota affiliates to set up shared pipelines, onboard their data, and support joint training and evaluation efforts.

Produce maintainable, well-documented code. Contribute to internal tooling and open-source releases to the scientific community.

Master’s or PhD in Computer Science, Electrical Engineering, Machine Learning, or a related field, with a minimum of 2 years of relevant experience and strong software engineering skills.

Deep proficiency in Python, PyTorch, and the Unix/Linux toolchain. Comfort working in terminal-heavy, SSH-based workflows on shared GPU clusters.

Hands-on experience with large-scale deep learning training, including distributed training (DDP, FSDP, DeepSpeed, or similar), GPU profiling, and debugging training failures at scale.

Experience building data pipelines for heterogeneous or multi-modal datasets (images, video, depth, point clouds, actions, etc).

Strong fundamentals in at least one of: computer vision, video understanding, generative models, 3D reconstruction, or robotics.

You are proactive, self-directed, and comfortable operating with ambiguity in a research-driven environment.

You are a reliable teammate who communicates clearly and takes ownership of problems end-to-end.

Experience with video diffusion models, world models, or multi-view geometry pipelines.

Familiarity with robotics data formats and collection pipelines (ROS, MCAP, HDF5, etc).

Experience with cloud training infrastructure (AWS SageMaker, EC2) and containerized workflows (Docker, Kubernetes).

Proficiency with modern AI-assisted development tools (e.g., Copilot, Cursor, Claude Code) for accelerating engineering workflows.

Track record of contributions to open-source projects or publications at top venues (CVPR, ICLR, NeurIPS, RSS, ICRA, etc.) is a plus but not required.