Multi-WAM World-v6 + Action-v6 reproducible handoff

This repository is the direct, auditable handoff for the Multi-WAM RoboTwin2.0 experiment. The latest snapshot is 2026-08-11: World-v6 main training is complete at step 97,314, its EMA branch is selected, the predicted-future cache is complete, Action-v6 has a portable step-10,000 checkpoint, and an intentionally compressed 22k-step Action continuation is running.

Start with PROGRESS_20260811.md for current progress, metrics and the documented Action protocol change. Exact identities, sizes and exclusions are in PACKAGE_MANIFEST.json and release-20260811/RELEASE_MANIFEST.json.

Latest portable state

Component	Included state	Purpose
Source	archives/source-code-action-v6-20260811.tar.zst	modified MIRA/LingBot source and experiment plans
World-v6	main-checkpoint-95000, main-checkpoint-97314	requested recovery point and selected final EMA source
Action-v6	action-checkpoint-10000	exact eight-rank restart boundary
Predicted futures	complete train/val/val_sample1 cache	required for Action steps after 10k
Evidence	metrics, fixed cases/videos, audits, logs and SwanLab local records	scientific traceability
Skill	skills/multi-wam-migrate/	restore, relocate, verify and eight-GPU resume workflow

The latest release tree is approximately 121.7 GB (decimal), of which about 67.7 GB is the checksum-addressed complete predicted-future cache. It deliberately excludes credentials, raw RoboTwin ZIPs, 503 GiB of reconstructable extracted RGB, 35 GiB of World anchor tensors unused by Action training, unselected recovery checkpoints, and two incomplete cache-probe directories.

The unchanged environment and base dependencies are reused from the previous release: the packed mira conda environment, runtime models, Qwen3-VL-2B, codec step-125000, World-80k migration source, and the 27,500-episode latent/text cache. No token or authentication store is included.

Direct selective download

Install a recent Hugging Face CLI and explicitly disable all proxies:

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 \
  --include README.md PROGRESS_20260811.md PACKAGE_MANIFEST.json SHA256SUMS \
    'skills/**' 'release-20260811/**' \
    archives/source-code-action-v6-20260811.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

The raw RoboTwin2.0 data is unchanged and intentionally external. Download or mount the 100 ALOHA ZIP files at workspace/datasets/RoboTwin2.0 before verification. Official dataset: https://huggingface.co/datasets/TianxingChen/RoboTwin2.0

Restore, verify and continue Action training

Read skills/multi-wam-migrate/SKILL.md, then run:

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_install.sh \
  /path/to/workspace /path/to/conda/envs/mira

bash skills/multi-wam-migrate/scripts/start_action_v6_resume.sh \
  /path/to/workspace /path/to/conda/envs/mira

restore_bundle.sh validates both checksum layers before extracting or copying payloads. verify_install.sh checks checkpoint shard counts and steps, model/cache identities, row and shard counts, source imports, tests, dataset presence and eight GPUs. The Action launcher clears every proxy variable, starts eight ranks in tmux, resumes a newer complete compressed checkpoint when one exists, otherwise starts exactly from Action-10k, and never embeds a credential in the repository.

The compressed Action schedule is 10k→13k ramp 0→0.5 predicted futures, 13k→16k at 0.5, 16k→19k ramp 0.5→0.75, and 19k→22k at 0.75. In-training validation is disabled by design; run the standalone exhaustive validation after step 22k before making final Action quality claims.

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.