Generative World Models Engineer – Robotics

Humanoid; Greater London; On-site; GBP 70,000 - 120,000; Full time

Benefits offered by this job

Stock options

30+ paid days off

Private healthcare

Pension scheme

Free daily breakfast and in-office食

In-office snacks

Direct access to leadership

Job summary

Humanoid is seeking a Research Engineer for the World Models team to build action-conditioned multimodal models predicting how our humanoid robots interact with real environments. You will design architectures, run large training jobs, and validate models against fleet data from industrial deployments.

You will work with diffusion models, transformers, and learned simulators to enable safe, scalable policy development and sim-to-real transfer in real-world robotics applications.

Qualifications

• Experience training large multimodal models (video, world, or multimodal).
• Deep hands-on experience with diffusion models, autoregressive transformers, or video prediction.
• Experience with large-scale distributed training and data pipelines.
• Strong Python + PyTorch/JAX; profile kernels and optimize data loaders.
• Empirical rigor in evaluating models and documenting experiments clearly.
• Interest in grounding generative models in physical reality, not pixels alone.

Responsibilities

• Design and train multimodal world models using diffusion-based and transformer architectures.
• Build action-conditioned video prediction and dynamics models with long-horizon physical consistency.
• Develop learned-simulator evaluation to predict real-world success and safety outcomes.
• Generate synthetic rollouts and counterfactual data to enhance policy training.
• Establish fidelity metrics and calibration protocols for world models.
• Build data pipelines turning fleet telemetry and video into training corpora.
• Run scaling and ablation studies on architecture, data mix, and context length.
• Collaborate with pretraining, RL, and manipulation teams to integrate world models into policy training.

Skills

Generative modeling

Python

Distributed training

Experimentation

Tools

PyTorch

JAX

Transformers

Job description

Humanoid is seeking a Research Engineer for the World Models team to build action-conditioned multimodal models predicting how our humanoid robots interact with real environments. You will design architectures, run large training jobs, and validate models against fleet data from industrial deployments.

You will work with diffusion models, transformers, and learned simulators to enable safe, scalable policy development and sim-to-real transfer in real-world robotics applications.