Founding Robot Learning Engineer | YC-Backed Physical AI & Robotics Startup | London (Remote) | Full-Time | Up to £150K + Equity

About the Company

Our client is a Y Combinator-backed (S26) physical AI data company building the foundational human video and motion datasets for leading robotics and embodied-AI teams. Focused on egocentric first-person video and UMI demonstration data, the company turns real-world capture operations into high-precision, versioned data products that foundation-model and VLA teams rely on to train physical robots.

Backed by a freshly closed seed round, their core constraint is supply, not demand. They are actively scaling deployment engineering, field operations, and capture hardware to keep up with massive customer demand.

The founders are building a three-person founding engineering team to set the technical standards, tooling, and evaluation methodology behind the world's reference dataset for robot learning.

About the Role

Reporting to the Engineering Lead (and working directly with partners and founders), the Founding Robot Learning Engineer bridges human demonstration data and real-world policy performance. You will decide, with empirical evidence, what datasets the company should collect and how that data should be represented.

This role balances hands-on data workflow engineering with policy experimentation. You will build task segmentation and assisted-labeling workflows, construct policy data loaders, and execute policy training and evaluation experiments using egocentric and UMI datasets to prove that the company's data directly improves robot task success.

Key Responsibilities

• Turn long recordings into useful task intervals, define annotation schemas for specific customer needs, and ensure every label remains frame-locked to the source video.
• Deploy AI models to propose task boundaries and labels, route uncertain results for human review, write labeling instructions, and measure reviewer consistency.
• Build custom data loaders and transformations that format video, labels, and motion trajectories (including human hand motion and UMI gripper paths) into policy-ready inputs.
• Fine-tune and evaluate robot policies on human demonstration data, running controlled baseline comparisons, held-out split tests, and task performance evaluations under realistic test conditions.
• Translate policy evaluation insights into precise data recommendations—guiding perception and capture teams on which tasks, viewpoints, and annotations yield the highest policy improvement.

Skills & Competencies

• Demonstrated experience training, fine-tuning, and evaluating robot manipulation or imitation learning policies (e.g., Diffusion Policies, VLAs, ACT).
• Strong Python and PyTorch capability with comfort building dataset interfaces, loaders, and annotation tooling alongside ML research.
• Deep understanding of evaluation leakage, dataset confounders, and the gap between offline loss metrics and physical robot task success.
• Experience working with temporal action segmentation, video annotations, spatial trajectories, and human-to-robot data transfer.

Qualifications & Experience

• Background as a Robot Learning Engineer, Embodied AI Researcher, or Robotics ML Engineer.
• Proven track record of personally preparing training data, training or fine-tuning robot policies, and evaluating physical or simulated task performance.
• Hands-on implementation experience connecting dataset decisions directly to policy improvements, identifying confounders, and preventing data leakage.
• Background in physical robot manipulation, diffusion policies, vision-language-action models, or human-to-robot transfer is highly desirable.

Role Details

• Timing: Immediate Start.
• Location: London (Remote).
• Package: Competitive base salary up to £150K + significant founding-engineer equity with absolute ownership over policy evaluation and dataset standards.