Skip to content

Add CPU actuator learning pilot and shoe skillspace plan - #8

Open
Tarzelf wants to merge 4 commits into
feat/motion-observation-v0.2from
feat/laptop-skill-pilot-v0.1
Open

Tarzelf wants to merge 4 commits into
feat/motion-observation-v0.2from
feat/laptop-skill-pilot-v0.1

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 11, 2026 •

Copy link
Copy Markdown
Contributor

What changes

Episode tapes now feed a compact, CPU-trained actuator transition model. The new dvidia-transition commands generate a grouped native corpus, fit and freeze models, score separate test groups, inspect retained artifacts, and produce an offline report. A shoe Skillspace capture plan defines the first bounded task and its missing evidence.

The retained experiment has 36 native recording attempts across 12 reset groups, split 8/2/2. The 6,438-byte action model predicts the next 20 ms of joint/jaw motion. Warm computation measured 0.207 ms at p95 on one ARM64 laptop, excluding sensing, transport and actuation. Commands reduce test joint/jaw RMSE; aggregate jaw MAE and worst error regress against persistence, and the first matched opening transition is substantially overpredicted. The report exposes these tradeoffs.

The predictor uses privileged simulator telemetry, emits no robot commands, and establishes no shoe policy, physical qualification or minimum laptop specification. Original learned coefficients and aggregate evidence are MIT licensed. Test transition targets are parsed only after model/protocol freezing; manifest terminal outcomes are visible during preflight.

Validation

  • 427 local tests passed, including 38 new core/pilot/report tests.
  • Actual measured report passed desktop and 390/320-pixel browser checks, slice/unit controls, exact receipt export, offline request checks and 44-pixel controls.
  • Built wheel installs cleanly; saved-model inference and evidence inspection run without NumPy or MuJoCo installed.
  • Copied public evidence bundle passes internal consistency inspection; no raw tapes, private paths, credentials or third-party weights included.
  • Shoe plan validates and accurately shows zero collected evidence and blocked task qualification.

Measured source: 0d3f05f186423e18b15ed8362d772b7d74ec0207. This candidate builds on the motion-observation branch (PR #7).

Windows CI exposed missing POSIX file flags and a system-default text encoding assumption. Portable regular-file admission, binary mode and explicit UTF-8 rendering now have regression coverage in a82ba9e. These loading/rendering fixes do not refit the models or alter the frozen evidence.

Report-test fixtures now write exact UTF-8 bytes; a CRLF/non-ASCII regression preserves exported receipt contents across Windows and Unix. Production inference and the retained measurements are unchanged.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant