Skip to content

CLEAR-LeWM v0.8.0: corrected runtime, audited reference

Latest

Choose a tag to compare

@DavidSunok DavidSunok released this 28 Aug 09:19
· 6 commits to main since this release

CLEAR-LeWM v0.8.0

CLEAR-LeWM v0.8 is a versioned evaluator correction and a fully audited RTX
4090 reference release. It does not rewrite or relabel the v0.5 evidence.

What changed

  • Reacher termination ownership. When CLEAR owns success scoring, v0.8
    disables the underlying dm-control task termination before rollout and
    reinstalls the gate after both reset() and model recompilation. This
    prevents an inner action_repeat step from emitting LAST and auto-resetting
    before CLEAR scores the external policy step.
  • Frozen comparison boundary. The 24 v0.8 manifests are byte-identical to
    v0.5. External Moderate/Strict predicates, fixed pairs, CEM settings, and
    checkpoint loading remain unchanged.
  • Audited reference matrix. The release includes 84 checksum-verified RTX
    4090 runs: official LeWM, paired random, DINOv2 No-Proprio LeWM, and GCBC
    Joint LeWM, using seeds 0/1/42 and 100 episodes per seed.
  • Updated public surface. README, GitHub Pages, synchronized task GIFs,
    result figures, and the corrected 1080p overview film now report v0.8.

Official LeWM results

Mean +/- sample standard deviation over seeds 0, 1, and 42:

Task Moderate Paired random Strict Paired random
PushT 86.67 +/- 0.58% 4.00% 70.67 +/- 5.51% 5.00%
Cube 50.33 +/- 3.79% 15.67% 21.67 +/- 4.73% 6.00%
Reacher 79.67 +/- 0.58% 7.33% 87.00 +/- 1.00% 8.00%
TwoRoom 83.00 +/- 5.29% 6.67% 51.33 +/- 3.21% 1.67%

All reference runs use pure CEM 300 x 30, top-k 30, solver batch size 1,
actor warm-start off, strict checkpoint loading, and the same manifest/policy
seed for model and random.

Reproducibility

Community credit

Thank you to @youzunzhi for identifying and
precisely isolating the Reacher action-repeat termination leak in
#2, and to
@08SCC for the unusually complete independent RTX
4090D reproduction and provenance audit in
#3. Both reports directly
shaped the version boundary and validation matrix for this release.

The immutable v0.5.0 and v0.5.1 tags/releases remain available. New work
should use v0.8; historical v0.5 bundles continue to be accepted only under
their explicit version label.