## About the role
Every robot session generates rich multimodal data, and its value depends entirely on the pipeline behind it. You will own the path from raw capture to training-ready datasets: automated episode triage, versioning and lineage, and the reproducibility discipline that makes training runs trustworthy.
## What you'll do
- Build an automated episode labelling layer: filter successes, failures and what needs human review
- Own dataset composition: coverage over task, object and environment space, near-duplicate removal, hard-example mining, balancing
- Run the human-in-the-loop review path
- Own reproducible training runs and regression gates
- Own the consumption side of the data pipeline, downstream of capture
- Perform data augmentation on the obtained datasets
- Work with the Physical AI and Navigation teams so the pipeline matches what they need
## What we look for
Some combination of the following
- You've built a data pipeline for a system fed by messy, real-world capture
- Dataset versioning and experiment tracking as daily working practice
- Enough perception literacy to build automated success/failure detection from raw sensor data
- Comfort with timeseries and multimodal data at scale
- Judgement about dataset composition, not just volume: you can say why a run improved and point at the data change behind it
## Tech stack
- Advanced experience with Python, Linux, Git
- PyTorch, and comfort reading training code you didn't write
- Docker, cloud object storage and infrastructure as code
- LeRobot / RLDS or comparable episode formats
- Dataset versioning and experiment tracking (DVC, MLflow, Weights & Biases or equivalent)
- Timeseries/multimodal tooling
## Bonus experience
- Active learning or DAgger-style correction loops
- Training-infrastructure cost optimisation
- Annotation tooling and vision models