huang2202/resop-runtime-assets-20261018

RoboTwin residual-RL runtime assets

Asset-only handoff published with the owner's permission on 2026-10-18. The project source code remains private. This repository does not contain the project implementation, credentials, offline demonstrations, feature caches, online replay, historical run trees, or upstream VLA model weights.

Minimum deployment set

Directory	Purpose
rlt/hanging_mug_332k/	Complete frozen single-task RLT snapshot: unchanged encoder.pt and manifest.json. Selected step 332000, complete-validation MSE 0.005848658366335763.
rlt/s4_2775k/	Complete frozen S4 pure-reconstruction best snapshot; selected step 2775000.
rlt/all50_1562500/	Frozen ALL50 pure-reconstruction best snapshot; selected step 1562500, from a run completed at 2500000 updates.
full50/no-clip-v4/	Complete eight-file Full50 v4 release, including the three seed registries, v9 action-scale map, manifest, attestation, README and license.
compatibility/	Exact normalization statistics plus pinned upstream model identities.
baselines/clean_base_e100_v1.json	Compact per-task frozen Base E100 summary: 50 tasks, 100 episodes each, 4500/5000 successes. Not an RLT/residual-policy evaluation.

Use an immutable commit revision, not a moving main, for deployments. The uploader supplies the final revision separately. No HF login is needed to download this public transport repository.

hf download huang2202/resop-runtime-assets-20261018 \
  --repo-type dataset --revision "$ASSET_REVISION" \
  --local-dir ./resop-runtime-assets
cd ./resop-runtime-assets
sha256sum -c SHA256SUMS

Set the existing runtime sources.selected_snapshot input to the absolute path of the desired snapshot directory. Keep encoder.pt and manifest.json together and unchanged. These assets initialize a frozen observation readout, not a pre-trained residual Actor/Critic or an online recovery state. No demonstration cache is needed simply to load the snapshot; a fresh experiment may still require its configured warm-up/collection stages.

The hanging-mug encoder SHA-256 / representation ID is 49123a7d9d9f257a71ffecd82b402cee5d31c43c2f6a02d1f8a84d1424b265e4. Validate with the project's existing load_frozen_readout(path, device="cpu") before allocating a training job. Loading must yield evaluation mode and zero trainable readout parameters.

Additional best-readout identities:

Snapshot	Encoder SHA-256	Best complete-validation MSE
S4	bb503450c9b9bdaefbf8cd22467fceff379e16c02b0723acbb291b75932b71bd	0.005648488186300186
ALL50	94f551733cf6e96bca6910a3cd2a0fe1953b7570b0ffab0f1db44d6600550f95	0.00318676792415965

These are reconstruction-selection statistics from different data/aggregation contracts; do not compare their absolute MSEs as a cross-task performance ranking or infer autonomous success rates. The ALL50 frozen pair was created with the existing consistent-best selector from best_full.json, complete validation events and encoder_best_full.pt, then loaded and verified on CPU. The encoder export bytes were preserved; only the deployment snapshot wrapper was newly materialized. The original 263 MB event log and full checkpoints are not shipped.

Full50 identity and relocation

This is a byte-for-byte transport mirror, not a new scientific release. The canonical release identity is still:

• Release: full50-no-clip-v4.
• Original Dataset: huang2202/resop-full50-no-clip-v4.
• Original revision: 68bbdd4045fc75645ceec6554c4a13ccb37caa71.
• Release manifest SHA-256: ac8e350db5c5d46615247b80ddbfd02e0bd31e94d13dba1166b2357587672007.
• Action-scale map: assets/action-scales/full50_action_scales_v9_no_clip.json, SHA-256 80907800417727b627973a9947249874776558a18b43e51c1ccf2d19f446ca10.

Point the existing local Full50-root input at full50/no-clip-v4 and use the tracked v4 descriptor/no-clip-v4 selector. Do not mix it with a v1 root or a different scale-map authority. The v9 map's 0.0005 values are ResFit-style action-scaling metadata, not an instruction to add action scaling to RLT full-action or base + linear correction experiments.

Do not run HF download with the Full50 subdirectory itself as --local-dir: the runtime root accepts only the release files (plus its optional install receipt), not .cache, a transfer README, or the parent handoff manifest. The layout above keeps HF metadata outside the strict release root. A normal v4 verifier can verify this local mirror without authenticating to the original private repository.

Absolute paths inside unchanged manifests are original-machine provenance; they are not portable destination paths. Do not edit those files to relocate them. Configure the new machine's asset/model locations using the existing runtime inputs. RLT producer identity also depends on the exact Base weights, normalization, preprocessing and selected feature layer: another compatible shape alone does not prove compatibility.

Upstream dependencies and licensing

Fetch public LingBotVLA/Qwen checkpoints and RoboTwin assets separately, using the pinned revisions in compatibility/upstream-models.json and the private project's environment instructions. Nothing here relicenses those upstream projects. Full50 retains its original CC-BY-4.0 license and attribution. Public availability of the RLT snapshots does not add a blanket license to the private project implementation or third-party model weights.

Full training-recovery checkpoints are deliberately separate: they can contain large replay and optimizer/RNG state, and must match the exact snapshot, configuration, phase and source revision. A bare Actor file is not a complete resume checkpoint.