VLA Pre-training & World Models | Humanoid

Role: VLA Pre-training Engineer / Deep Learning Engineer – World Models (2 openings) Company: Humanoid Location: London, UK (King’s Cross) – on-site, five days a week Job Type: Full-time About Humanoid Humanoid is building the world’s most capable, commercially scalable and safe humanoid robots. Its platform, HMND-01, is already deployed in real industrial environments, which makes Humanoid the most serious humanoid robotics company in Europe. 300+ people, 34+ nationalities, flat structure: your decisions ship directly into real robots. Offices in London (HQ, King’s Cross), Boston, Vancouver and San Diego. Values: Speed, Ambition, Passion, Trust. About the role The VLA Core team builds the base vision-language-action model that the entire manipulation pipeline depends on. This is deep learning at scale: training models from scratch on multi-embodiment trajectory corpora and internet-scale video, with real deployment constraints. The model has to fit and run on a real robot. Robotics experience is not required. If you come from LLMs, VLMs or video generation, be ready to get up to speed on a new domain quickly. This is a senior individual contributor role with no management track. The team is research-adjacent, but the output is models that run on robots. Results are shared as blog posts, not papers. There are two tracks in one team. They are the same role with a different focus, so you can lean into either. Track 1: VLA Pre-training Pre-train base models on a diverse multi-embodiment corpus of trajectories, fine-tune policies to excel at specific tasks and explore synthetic data. Post-train policies with behaviour cloning and RL, and own the full loop from data to deployment. Partner with the Data Collection team: specify what good data looks like, identify failure modes, ensure diversity and coverage. Run pre-, mid- and post-training on the VLA stack; explore new modalities and architecture changes. Build and maintain continuous pipelines: ingest synthetic data and teleop logs, version them, apply weak-supervision labelling, curate balanced datasets and automatically surface fresh failure cases into retraining. Work with the MLOps and Data Platform teams to scale distributed training and optimize models for real-time edge inference on the robot. Track 2: World Models Design and train multimodal, action-conditioned world models (video, state, action and language) using diffusion-based and transformer architectures. Build 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, roll out policy futures for safety review and planning. Generate synthetic rollouts and counterfactual experience (rare events, cross-platform and sim-to-real transfer) and measure their effect on downstream task performance. Build data pipelines that turn fleet telemetry, teleoperation logs and internet-scale video into training corpora. Run scaling and ablation studies on architecture, data mixture and context length. What you need to succeed 3+ years building deep learning systems in industry or research, with shipped models or published work to show for it. Deep hands-on experience with at least one of: LLMs, VLMs, image or video generative models (architecture, training and inference). For the World Models track: diffusion models, autoregressive transformers, latent-variable models or video prediction. You have actually trained models yourself: pre-training runs, distributed infrastructure, the works. Connecting APIs, building agents or a quick fine-tune is not enough. Experience with deep learning infrastructure: streaming datasets, checkpointing and state management, distributed training strategies, debugging numerics and training instabilities. Strong