How Two GitHub Issues Shaped RDA v0.9.14: Action Units, Per-Joint Analysis, and Smarter Freeze Detection

Description and use

Two GitHub issues led to RDA v0.9.14, which now correctly infers action units, exposes per‑joint spike counts, guarantees scale invariance, and adds state‑based freeze cross‑validation.

Problem

Robot learning datasets often contain action data in inconsistent units, making spike detection unreliable and hindering dataset quality assessment.

Outcome

RDA v0.9.14 improves data quality diagnostics by correctly handling action units, exposing granular per‑joint spike data, ensuring scale invariance, and enabling state‑based freeze detection, all driven by real user feedback.

Workflow

• 1. User files GitHub issues with reproducible analyses.
• 2. RDA developers review issues, run tests, and confirm findings.
• 3. RDA code is updated to infer action units, expose per‑joint data, add scale‑invariance guard test, and implement freeze‑motion‑source flag.
• 4. New version is released on PyPI and GitHub.
• 5. Users run RDA v0.9.14 on their datasets and interpret updated reports.

Evidence

We ran the test properly. We took the same svla action data and applied four different scale transforms… The result: spike_count was 260 in all four cases.

Source

In v0.9.14, we moved by_joint to the top-level measurement output… Running v0.9.14 on the full svla dataset gives per-joint totals… very close to the developer's independent calculation.

Source