Deep Learning Engineer - World Models

About the role

Here at Humanoid, we believe in a future where robots amplify human potential. That’s why we’ve set out on a mission to build the world’s most capable, commercially-scalable, and safe humanoid robots. We’re bringing that mission to life with HMND‑01 - our rapidly developed humanoid platform being deployed in real industrial environments - and we’re growing the team to take it even further.

OUR MISSION

At Humanoid we strive to create the world's leading, commercially scalable, safe, and advanced humanoid robots that seamlessly integrate into daily life and amplify human capacity.

ABOUT THE ROLE

As a Research Engineer on the World Models team, you will build action-conditioned generative models that predict how the world evolves around our robots — future video, proprioception, contacts, and outcomes — from past observations and actions. World models serve four purposes in our stack: a pretrained, physics-aware prior for our VLA policies; an engine for rare data collection, cross-platform transfer, and sim-to-real transfer; a testbed for policy evaluation and testing before hardware; and a future-prediction rollout engine that surfaces what our policies intend to do, for safety and planning. This is a hands-on individual contributor role: you will design architectures, run large training jobs, and validate your models against real fleet data from industrial deployments.

WHAT YOU'LL DO

- Design and train multimodal world models — video, state, action, and language — using diffusion-based and transformer architectures.

- Build action-conditioned video prediction and dynamics models that stay physically consistent over long horizons, including contact-rich manipulation, and serve as pretrained priors for VLA policies.

- Develop learned-simulator evaluation: score candidate policies offline, predict real-world success rates before deployment, and roll out policy futures to expose intended behaviour for safety review and planning.

- Generate synthetic rollouts and counterfactual experience — including rare events, cross-platform transfer, and sim-to-real transfer — to augment policy training, and measure their effect on downstream task performance.

- Establish fidelity metrics and calibration protocols that quantify where the world model can be trusted and where it diverges from reality.

- Build data pipelines that turn fleet telemetry, teleoperation logs, and internet-scale video into training corpora for world models.

- Run scaling and ablation studies on architecture, data mixture, and context length; communicate findings crisply.

- Collaborate with pretraining, RL, and manipulation teams to integrate world models into policy training and evaluation loops.

WHAT WE'RE LOOKING FOR

- A track record of training large generative models — video, world, or multimodal — with shipped models or published artifacts to show for it.

- Deep hands-on experience with modern generative architectures: diffusion models, autoregressive transformers, latent-variable models, or video prediction.