Multi-WAM reproducible experiment handoff

This repository is the checksum-audited handoff of the Multi-WAM RoboTwin2.0 experiment. The current full snapshot is 2026-09-02. World-v12 training is complete at step 25,066, the final Action-v13 continuation is complete at step 23,016, and the exact 307 GiB Action predicted-future cache is included as a checksum-audited overlay.

Component	Packaged checkpoint	Status
World-v12	release-20260902/workspace/outputs/robotwin_bihub_world_v12_largebatch_remaining/main-checkpoint-25066	final training recovery point
Action-v12	release-20260902/workspace/outputs/robotwin_bihub_wam_v12_action_eval_aligned_epoch/action-checkpoint-5754	one-epoch lineage parent
Action-v13 best offline	release-20260902/workspace/outputs/robotwin_bihub_wam_v13_action_eval_aligned_continue4epoch/action-checkpoint-11508	best fixed-panel MSE
Action-v13 final	release-20260902/workspace/outputs/robotwin_bihub_wam_v13_action_eval_aligned_continue4epoch/action-checkpoint-23016	final closed-loop eval weight
Codec	archives/codec-checkpoint-step125000.tar.zst	required, already Hub-resident

Read PROGRESS_20260902.md for the training design, lineage and metric definitions, EVALUATION_20260902.md for current RoboTwin results and their limitations, and REPRODUCE_20260902.md for exact migration steps.

The completed 8-task clean/unseen probe scores 29/64 = 45.31%. A later clean/seen 50-task run was interrupted after 25 tasks and 1,052 valid episodes; its 586/1,052 = 55.70% is explicitly a partial observation, not a formal 50-task score.

Direct selective download

Use the pinned snapshot and a direct connection:

unset http_proxy https_proxy all_proxy ftp_proxy HTTP_PROXY HTTPS_PROXY ALL_PROXY FTP_PROXY
export NO_PROXY='*' no_proxy='*' HF_XET_HIGH_PERFORMANCE=1

hf download Orangerl/multi-wam --revision snapshot-20260902-full \
  --include README.md PACKAGE_MANIFEST.json SHA256SUMS \
    PROGRESS_20260902.md EVALUATION_20260902.md REPRODUCE_20260902.md \
    'skills/**' 'release-20260811/**' 'release-20260814/**' \
    'release-20260829/**' 'release-20260902/**' 'release-20260902-action-cache/**' \
    archives/source-code-world-v12-action-v13-20260902.tar.zst \
    archives/eval-code-robotwin-v13-20260902.tar.zst \
    archives/mira-conda-env.tar.zst archives/runtime-models.tar.zst \
    archives/qwen3-vl-2b-instruct.tar.zst archives/codec-checkpoint-step125000.tar.zst \
    archives/world-checkpoint-step80000.tar.zst archives/world-derived-cache.tar.zst \
    archives/world-v6-migration-inputs.tar.zst \
  --local-dir multi-wam-hf

Then restore and verify:

cd multi-wam-hf
bash skills/multi-wam-migrate/scripts/restore_bundle.sh \
  "$PWD" /path/to/workspace /path/to/conda/envs/mira
bash skills/multi-wam-migrate/scripts/verify_current_release.sh \
  /path/to/workspace /path/to/conda/envs/mira

The release includes current training code, eval code, checkpoint states, fixed cases/videos, offline validation metrics, raw closed-loop episode diagnostics, important logs, and the exact Action-v12/v13 predicted-future cache (329,522,409,844 bytes). Raw/extracted RoboTwin data and the about 3.1 TB World Qwen-hidden cache remain external. Credentials are never part of the current source/release payload. For inference/eval-only downloads, the Action cache layer may be omitted; the full restore command intentionally requires it.

MIRA and LingBot-Vision retain their upstream Apache-2.0 notices. Bundled third-party models, checkpoints, data derivatives and conda packages retain their respective licenses, so this aggregate repository uses the mixed/other label rather than relicensing those assets.