Member of Technical Staff - World Models

Stealth AI Startup; Redwood City (CA); On-site; USD 220,000 - 280,000; Full time

Job summary

Stealth AI Startup in Redwood City seeks a Founding Member of Technical Staff to advance how robot policies are learned from human video. You will shape representations of actions and align video objectives with learned actions, integrating post-training against real robot behavior.

The team hails from Meta, Google, and CMU, with multi-million seed funding. This is an in-person role in Redwood City offering founding-tier equity and visa sponsorship.

Qualifications

• Trained a model that outputs actions at scale.
• Hands-on with video generative models at scale.
• Experience with learning actions from unlabeled video or VLA/post-training.
• Released code/checkpoints or a first-author paper used by others.
• Led a full training run from data to infra to evaluation.

Responsibilities

• Work on how actions are represented and learned from human data, and how video objectives train with action models; develop post-training against robot behavior.

Skills

Model training at scale

Video generative models

Learning from unlabeled video

VLA training or post-training

Diffusion or flow-matching policies

Job description

We scale dexterous robot policies by learning from human video. Our team of researchers and engineers comes from Meta, Google, and CMU, with a decade of work on Computer Vision, Computer Graphics, Foundation LLMs, and Robotics. Double-digit millions raised, led by a tier-one seed fund. In person in Redwood City.

Role Description

You work on how actions are represented and learned from human data, how the video and action objectives train together, and how the model gets post-trained against what a robot actually did. Title is Founding Member of Technical Staff.

Compute is a cluster of B300s. Cash is $220K to $280K depending on experience, plus founding-tier equity. O-1 and H-1B sponsorship.

Qualifications

• You have trained a model that outputs actions at scale
• Hands-on with at least one of: video generative models at scale, learning actions from unlabeled video, VLA training or post-training, diffusion or flow-matching policies.
• You have hit the embodiment gap in practice and can say concretely where it breaks.
• Released code, checkpoints, or a first-author paper people build on.
• You have carried a full training run alone, from data to infra to training to evaluation.