ML Research Engineer, Data

Weave Robotics•San Francisco, CA

About The Position

Weave builds robots for homes and businesses and evaluates robotics research by whether robots can perform useful work in real-world settings. Its fleet generates terabytes of data that feed weekly training runs. The ML Research Engineer, Data will architect and build the data pipeline that turns raw fleet data—including video, proprioception, actions, and sensor streams—into datasets for model training, and will help shape data sampling ratios, mixtures, and curricula with the research team. The role combines data engineering with analytical understanding of the data. It involves building a platform to process millions of episodes and serve data to training at full throughput, identifying data characteristics associated with model behavior, and creating targeted datasets and evaluation benchmarks to improve and measure models across tasks, environments, embodiments, and model versions.

Requirements

  • Experience building data engineering pipelines over terabytes, using object stores such as S3 or GCS, distributed storage, and indexing.
  • Experience with high-throughput training formats and dataloaders.
  • Working ML experience, including launching a fine-tune, interpreting a loss curve, and designing an ablation to test a data hypothesis.
  • Ability to apply signal processing and statistics to real sensor data and use unsupervised methods such as PCA, UMAP, and clustering when labels are unavailable.
  • Ability to debug data problems across sensor drift, corrupted episodes, and pipeline failures.
  • Strong Python skills and software engineering fundamentals.

Nice To Haves

  • C++ experience.
  • Experience with batch processing at scale using Ray, Spark, or Dask, including cost and throughput judgment.
  • Experience with production workflow orchestration such as Airflow, Kubeflow, or similar, including retries and monitoring.
  • Knowledge of robotics data pitfalls, including timestamps, clock domains, and sensor modality quirks.
  • Robot learning experience, such as training policies using VLAs, world models, or RL, and distinguishing data problems from model problems.

Responsibilities

  • Characterize data coverage, redundancy, and drift using time-series spectral analysis, distributional statistics, and clustering over embeddings.
  • Identify bad data in large episode collections, including bad trajectories, bad annotations, dropped frames, and desynchronized streams, and automate detection to prevent it from entering the pipeline.
  • Build the data lifecycle, including curation, preprocessing, annotation, augmentation, versioning, training-ingest formats, and loaders.
  • Use embedding search, model loss and disagreement, and vision-language-model-assisted filtering and labeling to analyze and curate data.
  • Partner with researchers to investigate model failures and build datasets, processing steps, and samplers targeting specific capabilities and failure modes.
  • Own data sampling and mixture decisions for training runs.
  • Build evaluation datasets and benchmarks measuring performance across tasks, environments, embodiments, and model versions.
© 2026 Teal Labs, Inc
Privacy PolicyTerms of Service