VLA Pre-training Lead (Deep Learning)

Job description

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 Alpha - our rapidly developed humanoid platform now running in real industrial pilots - 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 Pretraining VLA Lead, you will own the pretraining stage of our VLM and VLA-based framework powering the fleet across wheeled and bipedal platforms. You will define the base model strategy — architecture, data mixture, and scaling — and lead a team of research engineers training foundation policies on a diverse, multi-embodiment corpus of real-world trajectories, teleoperation and synthetic data, and internet-scale video and language data. The base models you deliver are the substrate every downstream team fine-tunes for locomanipulation, so your decisions shape the capability ceiling of every robot we ship. # What You'll Do - Own the VLA pretraining roadmap end-to-end: architecture choices, data mixtures, scaling laws, and evaluation protocols for base models. - Push pretraining beyond a single recipe: explore transformer- and diffusion-based architectures, video pretraining, and world-model objectives that turn multimodal data (video, action, state, language) into generalisable robot capabilities. - Lead, grow, and mentor a team of deep learning engineers focused on pretraining, setting research direction and engineering standards. - Design and run large-scale distributed training on multi-node GPU clusters; drive throughput, stability, and cost efficiency in partnership with MLOps & Data Platform teams. - Define what pretraining-scale data looks like: partner with the Data Collection team and external data providers to secure a steady supply of high-quality, diverse, multi-embodiment trajectories. - Build rigorous base-model evaluation suites that predict downstream post-training and real-robot performance, and use them to make principled go/no-go scaling decisions. - Establish continuous pretraining pipelines: dataset versioning, curation, deduplication, weak-supervision labelling, and automatic surfacing of coverage gaps. - Collaborate with post-training and RL teams to ensure base models transfer cleanly to fine-tuning and real-time edge inference. - Track and drive the frontier: evaluate emerging VLA architectures, modalities, and training recipes, and decide what enters our production stack. # What We're Looking For - A track record of building deep-learning systems (industry or research), with shipped models or published artifacts to show for it, and experience leading a team or a major workstream. - Proven experience pretraining large models — LLMs, VLMs, video/generative models, or VLAs — at multi-node scale: you have owned data mixtures, scaling decisions, and training stability for large distributed runs. - Deep understanding of transformer and diffusion architectures, multimodal training, and the practicalities of distributed training - Strong Python + PyTorch/JAX; you can debug and profile ML systems and write maintainable research code. - A track record of making data-driven scaling decisions and communicating trade-offs crisply to both researchers and leadership. - You document experiments clearly and build teams that do the same. ## Nice to have - Experience with VLA (vision-language-action) models and frameworks. - Robotics or autonomous driving experience, especially multi-embodiment or cross-platform learning. - Experience with synthetic data generation and sim-to-real pipelines at scale. - Publications at top-tier deep learning conferences (NeurIPS, ICML, ICLR, CoRL) or equivalent open-source contributions. - Experience optimising foundation models for real-time edge inference. # What We Offer - Competitive equity: stock options with meaningful upside as we scale. - 30+ paid days off, including 23 days of annual leave, all UK bank holidays, and additional company closure days (including Christmas–New Year shutdown). - Private healthcare, including virtual and in-person care. - Pension scheme with 8% total contribution (5% employee, 3% employer) on full earnings. - Free daily breakfast, catered lunch, and snacks in-office. - Work at the frontier - collaborate daily with world-class engineers, researchers, and product experts building the next generation of AI and humanoid robotics. - Real ownership - direct access to founding leadership, meaningful input on product direction, and the ability to drive key initiatives from day one.