ML Research Engineer, Data

Weave Robotics • San Francisco, CA

1d

About The Position

Most robot learning research is graded on evals that don't survive contact with the field. Ours is graded by robots doing useful work in real homes and businesses, every day. Our fleet generates robot data at terabyte scale that’s ingested by training runs every week. Our models are only as good as what we feed them, and you’ll architect and build the pipeline that drives their behavior.

You'll turn raw fleet data (video, proprioception, actions, sensor streams) into the datasets our models train on, and follow that data into the training loop: sampling ratios, data mixes, and curricula are decisions you'll shape with the research team. The job is equal parts data engineering and data understanding: build the platform that processes millions of episodes and feeds them to training, and know the data well enough to say what correlates with good and bad model behavior.

Requirements

• Data engineering at scale: pipelines over terabytes, object stores (S3, GCS), distributed storage, and indexing.
• Training ingest: high-throughput formats and dataloaders.
• Working ML experience: you can launch a fine-tune, read a loss curve, and design an ablation to test a data hypothesis.
• Analytical range: signal processing and statistics on real sensor data, and unsupervised structure-finding (PCA, UMAP, clustering) when the labels don't exist yet.
• Data debugging: you can trace a problem from sensor drift through a corrupted episode to a pipeline failure.
• Strong Python and software engineering fundamentals.
• C++ is a plus.

Nice To Haves

• Batch processing at scale: Ray, Spark, or Dask over terabytes, with cost and throughput judgment.
• Training ingest: formats and dataloaders that keep GPU clusters fed.
• Workflow orchestration in production: Airflow, Kubeflow, or similar, with retries and monitoring.
• Robotics data pitfalls: timestamps, clock domains, and sensor modality quirks.
• Robot learning exposure: you’ve trained policies (VLAs, world models, RL) and can tell a data problem from a model problem.

Responsibilities

• Know the data: characterize coverage, redundancy, and drift: spectral analysis on time-series, distributional statistics, clustering over embeddings.
• Find the needles: dig bad data out of terabytes of episodes: bad trajectories, bad annotations, dropped frames, desynced streams, then automate the catch so it never gets through again.
• Build the data lifecycle: curation, preprocessing, annotation, augmentation, versioning, and the training-ingest formats and loaders that serve it at full throughput.
• Use models as instruments: embedding search to mine scenarios, model loss and disagreement as quality signals, VLM-assisted filtering and labeling.
• Improve models through data: partner with researchers on model failures, then build the datasets, processing steps and samplers that target specific capabilities and failure modes, and own the sampling and mixture decisions that go into each run.
• Build the eval datasets and benchmarks that measure performance across tasks, environments, embodiments, and model versions.

What This Job Offers

• Job Type: Full-time
• Career Level: Senior
• Education Level: No Education Listed