Skip to content
Jangara Bliss
All projects
Research & InvestigationIn progress2026 · foundation complete

Physical AI Data Flywheel

A closed-loop data-to-policy system that turns robot benchmark failures into targeted demonstration requests — then validates, trains, evaluates, and repeats.

Role — Creator — system design, schemas, validation, selection strategies, artifact pipeline

Pythonxembench (ManiSkill3)Typed dataset / provenance schemasHDF5 validationFailure-driven candidate generationArtifact packaging
Problem
Robot-learning pipelines usually treat data collection as a fixed upfront cost: gather demonstrations, train, hope. But when a policy fails, those failures say exactly where the dataset is thin. The flywheel asks whether closing that loop — failures deciding what data comes next — beats collecting more of the same.
System type
Data-to-policy loop · dataset validation & selection infrastructure
Why it matters
Data quality and data choice are becoming the bottleneck of physical AI. A trustworthy flywheel needs validation, provenance, and honest accounting built in from the start — otherwise 'more data' quietly becomes 'more of the same failure modes'.
Team context
Solo project, built as a layer inside the x-embodiment benchmark codebase.
Physical AI data flywheel loop: evaluation failures flow into failure analysis, candidate demonstration requests, budgeted selection, collection and dataset validation, policy training and re-evaluation, and round artifact packs — which feed the next round
One flywheel round — failures decide what data gets collected next.

01

Overview

Robot demonstrations are expensive, so which ones you collect matters. Instead of gathering more data at random, this flywheel — built inside the cross-embodiment benchmark — reads evaluation failures, generates candidate demonstration requests aimed at what actually failed, selects among them under an episode budget, validates the resulting dataset before any training touches it, and packages every round into a reproducible artifact with dataset cards and provenance. The Mac-safe foundation is complete: seventeen modules, typed schemas end to end, deterministic validation, five selection strategies, and a full toy round rehearsed on synthetic data with explicit warnings. The real collect–train–evaluate loop is pending Linux experiments, which are designed to test one falsifiable question: does failure-targeted data improve performance more efficiently than the same budget of random additional demonstrations?

System architecture

Evaluation episodes stream failure records into analysis; failures become candidate demonstration requests; a selection strategy spends a fixed episode budget across candidates; collected demonstrations pass deterministic validation before entering the training set; the retrained policy is re-evaluated, and the whole round — dataset card, validation report, plans, and comparisons — is packaged as a checksummed artifact. Then the loop runs again.

  1. Evaluation failures → failure analysis
  2. Candidate demonstration requests
  3. Budgeted selection — 5 strategies
  4. Collection → dataset validation
  5. Policy training → re-evaluation
  6. Round artifact pack — cards · provenance

02

Contributions

  • Designed the round loop: evaluation failures → failure analysis → candidate demonstration requests → budgeted selection → collection → dataset validation → training → re-evaluation → artifact packs.
  • Built seventeen flywheel modules with typed schemas for provenance, dataset index records, quality issues, candidate demo requests, selection plans, dataset cards, model-card placeholders, and round manifests.
  • Implemented deterministic dataset validation with quality scoring — datasets are checked before any training consumes them.
  • Implemented five budgeted selection strategies — random, stratified, failure_targeted, diversity_greedy, and cost_aware_failure_targeted — so the core comparison (targeted vs. random) is a config switch, not a rewrite.
  • Rehearsed the entire loop as a deterministic toy round producing synthetic artifact packs with explicit warnings, and generated Linux command plans without executing any training on the Mac.
  • Covered the layer with 77 flywheel tests; the repo-wide suite stands at 187 fast + 5 simulation tests.

03

Evidence & evaluation

Evidence

Flywheel foundation

attached

17 modules, typed schemas end to end, deterministic validation with quality scoring, and failure-driven candidate generation.

Selection strategies

attached

random · stratified · failure_targeted · diversity_greedy · cost_aware_failure_targeted — all budgeted, all tested.

Toy round rehearsal

attached

Deterministic end-to-end round on synthetic data, packaged into artifact packs with explicit synthetic-data warnings.

Test suite

attached

77 flywheel tests; repo-wide suite at 187 fast + 5 simulation tests, all passing.

Real flywheel rounds

pending

The collect–train–evaluate loop on real data awaits Linux experiments; all current outputs are synthetic/toy.

Targeted-vs-random experiment

pending

Pre-registered comparison: baseline vs. +random demos vs. +failure-targeted demos at equal budget.

Metrics

Improvement per demonstration

Not yet measured

The flywheel's headline metric; pending real Linux rounds.

Failure-targeted vs. random selection

Not yet measured

Equal-budget comparison; pending real experiments.

Dataset quality score

Not yet measured

Computed by validation on real collected data; pending.

Flywheel modules

17

Flywheel tests

77

04

Limitations

  • Mac-safe foundation complete; the real data/training/evaluation loop is pending Linux experiments — no flywheel results exist yet.
  • All current flywheel outputs are synthetic/toy and are explicitly labeled as such unless marked real later.
  • Failure targeting currently works at the granularity of the benchmark's existing failure labels; sharper targeting lands with richer per-episode instrumentation.
  • Designed and rehearsed on one benchmark (xembench); generality beyond it is a hypothesis, not a claim.

05

Lessons & tradeoffs

  • Validation belongs before training, not after a bad run: a dataset the pipeline can't trust is a result you can't trust.
  • Making 'which data next?' a typed, budgeted, testable decision turns a vague intuition — failures should guide collection — into an experiment with a falsifiable answer.

06

Artifacts

  • Code (private — public release pending)not yet published
  • Flywheel experiment reportnot yet published
  • First real round artifact packnot yet published