Skip to content

[Robot] restore station rollout knobs on the dashboard - #319

Draft
aidang3019 wants to merge 1 commit into
codex/arc-origin/projection-progressfrom
aidan/rollout-knobs-20261025
Draft

aidang3019 wants to merge 1 commit into
codex/arc-origin/projection-progressfrom
aidan/rollout-knobs-20261025

Conversation

@aidang3019

Copy link
Copy Markdown

Stacked on #317. Brings back the rollout knobs the Tuesday (#263) dashboard had. #317's ARC profile shows only sampling steps, and it opens at 10.

  • Station-owned controls (policy.station_controls in the three YAM rollout profiles): Actions before replanning, Fastest-stream termination, Execute first % of waypoints, Replay speed, Hold speed. A control is added only when the model declares none of that name and, for decoder controls, only when the loaded decoder has the attribute. Time bundles keep their own controls (the station adds nothing to them). Checkpoint-bound declarations are untouched, so bundles already exported get the knobs as is.
  • policy.inference_defaults: sampling steps (Euler steps) open at 50, the trained and validated value, for any model that exposes the control.
  • ARC replay tempo: FirstStreamArcDecoder reuses the ReplayTempo warp (moving vs hold speed) on the translation clock. Decode-only codec helpers: detokenize(times=), stream_clocks, control_duration.

Checks (CPU, WSL):

  • At 100 % / 100 % the decode is byte-identical to 776656e on 60 real val chunks, duration/translation and velocity/full_separation, M=200.
  • At 2×, row k equals 1× row 2k+1 exactly and the fastest-stream step count is ceil(1×/2).
  • Hold 100 % with replay 200 % lands between the two.
  • audit_hydra_configs: 325/325. audit_components: 276/325, the same failure set as 776656e (OAT / LIBERO / Qwen / egobridge env-only).

Not yet run on the arms.

The PR #317 ARC contract exposes only sampling steps (opening at 10). Add
station-owned controls in the YAM rollout profiles, bound only where the model
declares none of that name and the loaded decoder has the target:
Actions before replanning, Fastest-stream termination, Execute first % of
waypoints, Replay speed and Hold speed. `policy.inference_defaults` opens
sampling steps at the trained 50. Checkpoint-bound declarations are unchanged,
so already-exported bundles get the knobs without re-export.

FirstStreamArcDecoder gains the ReplayTempo warp (moving vs hold speed) on its
translation clock; the codec gains decode-only `detokenize(times=)`,
`stream_clocks` and `control_duration`. At 100 % / 100 % the decode is
byte-identical to 776656e.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSaa5mnstFBVJi8329KLDc
This was referenced Oct 26, 2026

aidang3019 commented Oct 26, 2026 •

Copy link
Copy Markdown
Author

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

This stack of pull requests is managed by Graphite. Learn more about stacking.

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