Vincent Deng

Work portfolio
github.com/vincentstanfordlead

Hands-on with robot data, from teleop collection to QA and real-robot eval. Ran large-scale data programs with external vendors, from spec and contracting to delivery and QA.

Recent builds

  1. Cross-embodiment real-robot eval

    EvalCosmos ReasonPiperxArmYAM

    Ran 4 frontier models zero-shot on 3 different arms, with Cosmos Reason 2 judging each run alongside human graders. Built the eval pipeline, a mobile launcher and the results dashboard.

    Repo

  2. Is public robot data GR00T-ready?

    Data standardsQANVIDIA GR00TLeRobotHugging Face

    Checked 562 public LeRobot datasets against GR00T’s data loader: 30.5% are one file away and 58% need a format conversion. Built a website so anyone can check their own data’s readiness.

    Web appRepo
  3. Volume funnel and time funnel, illustrative numbers Volume, hours of footage indexed to 100: uploaded 100, review started 90, reviewed 72, idle removed 57, effective hours 50. Time, median hours per stage from recording to QA verdict: waiting to upload 55, uploading 3, ingest 9, queue for review 32, human QA 24. Waiting to upload is the biggest delay. VOLUME · HOURS TIME · MEDIAN HOURS Uploaded 100 Review started 90 Reviewed 72 Idle removed 57 Effective hours 50 Waiting to upload 55 h Uploading 3 h Ingest 9 h Queue for review 32 h Human QA 24 h ILLUSTRATIVE NUMBERS

    Egocentric data collection funnel

    Egocentric dataCollection opsQAFunnel analytics

    Standardized the funnel definition for a new egocentric collection channel and found where usable data was being lost. Built an ops dashboard that turned findings into feedback for data collectors to raise QA pass rate.

  4. 6-DOF arm teleoperation pipeline

    TeleoperationData collectionQAPolicy trainingEval

    Assembled a 6-DOF arm and ran the data pipeline end to end: teleop, collection, QA, policy training and eval. 37 demos surfaced 5 failure modes, most traced back to the data.

    Repo
Data programs at scale
  1. Program timeline from July to February: TransUnion contracting and data agreement, survey and panel development, survey deployment, data matching and listening period, then analysis and report, with milestones from proposal and funding to executive report

    PII data matching program

    Data acquisitionPII matchingTransUnionLegal & privacy reviewQA

    Linked brand perception to consumer behavior by matching third-party survey data with first-party transaction data at the user level: 6 months, 4 companies, 35 stakeholders, millions of customer records.

  2. Aerial rendering of Intuit Dome with the roof logo highlighted

    Intuit Dome sponsorship impression data

    Data acquisitionVendor QANielsenLA Clippers

    Measured billions of impressions across TV, social, news, PR and on-site channels to value a $600M naming-rights deal, scoring each impression by quality, not just volume.

  3. Answer distributions from a human panel next to two AI synthetic user panels

    AI synthetic user simulation

    Synthetic dataModel evalHuman calibration

    Built 600 synthetic users that answer product surveys like small-business owners, since research panels are slow and expensive. On most questions, AI answers closely tracked a large human panel.