Research Engineer/Scientist, Imitation Learning & Pretraining Research · San Francisco · Full-time →

Tau Robotics, Inc.; San Francisco (CA); On-site; USD 140,000 - 200,000; Full time

Job summary

Tau Robotics, Inc. in San Francisco seeks a Research Engineer/Scientist focused on imitation learning and pretraining to advance robot autonomy.

You will design and scale pretraining pipelines, train multi-task manipulation policies, and define finetuning methods integrated with reinforcement learning. This role emphasizes hands-on experimentation with large multimodal datasets and strong Python + PyTorch or JAX experience, with meaningful equity.

Qualifications

• Experience with large-scale pretraining, behaviour cloning, or vision-language-action models.
• Comfort building data pipelines for large multimodal datasets.
• Strong Python plus PyTorch or JAX.
• Evidence: papers, repos, or models you trained that others used.

Responsibilities

• Build and scale supervised pretraining runs on large multi-modal robot datasets.
• Train multi-task manipulation policies and study what actually transfers.
• Define the finetuning recipes that hand pretrained policies to the RL loop.

Skills

Python

PyTorch

JAX

Multimodal datasets

Vision-language-action models

Job description

Research Engineer/Scientist, Imitation Learning & Pretraining.

Our mission is to build a general-purpose AI for robots. We believe the path is to deploy robots early in real customer environments, learn from large-scale real-world experience, and steadily increase autonomy and performance beyond human levels. Join a focused team building this system end to end, with significant ownership and meaningful equity.

What you’ll do

• Build and scale supervised pretraining runs on large multi-modal robot datasets
• Train multi-task manipulation policies and study what actually transfers
• Define the finetuning recipes that hand pretrained policies to the RL loop

What we look for

• Experience with large-scale pretraining, behaviour cloning, or vision-language-action models
• Comfort building data pipelines for large multimodal datasets
• Strong Python plus PyTorch or JAX
• Evidence: papers, repos, or models you trained that others used