From retained records to published claims
This page identifies the artifacts and commands behind the numerical claims on the project site. It documents the completed work and does not authorize new training runs or robot trials.
What can be reproduced
Values derived from data are recomputed from retained CSV, Parquet, JSON, and annotation files. Configuration values are checked against policy code, resolved commands, checkpoint configuration, and evaluation manifests. Literature claims are checked against same-key source files in the reference archive.
Collection and calibration lineage
| Component | Authoritative record | Verified property |
|---|---|---|
| Peg-in-hole dataset | experiments/datasets/peg-tight-vertical-100/ | Episode count, frame count, timestamps, camera keys, states, and actions |
| Controlled views | act-v/meta/ and act-m/meta/ | Identical episode and frame identities with only the state view changed |
| Leader calibration | calibration/so101_leader.json | Joint order, servo identifiers, ranges, homing offsets, and digest |
| Follower calibration | calibration/so101_follower.json | The same fields and digest, fixed from collection through evaluation |
Identity before interpretation
Resolved command records preserve dataset view, seed, update budget, policy family, resolution, and chunk or horizon settings. Evaluation manifests bind deployed labels to checkpoint hashes and policy-code revisions. Trial rows preserve order, outcome, failure phase, and session context.
Deterministic checks
python3 research/writing/audit_final_manuscript.pypython3 research/writing/verify_chapter5_dataset.pypython3 research/writing/verify_chapter7_evaluation.pyLarge recordings and checkpoint tensors remain in the project archive and are not embedded in this website.