multi-wam / README.md
Orangerl's picture
Add files using upload-large-folder tool
d9ded1c verified
|
Raw
History Blame Contribute Delete
3.77 kB
metadata
license: other
tags:
  - robotics
  - world-model
  - robotwin
  - multi-view
  - pytorch

Multi-WAM World v6 migration bundle

This public repository is a reproducible transfer bundle for the current Multi-WAM RoboTwin2.0 World Model v6 stage. The 2026-08-05 update adds the complete modified v6 source, frozen Qwen3-VL-2B runtime, the audited World step-80000 migration checkpoint, migration inputs, deployment evidence, and an agent Skill. Unchanged, SHA-identical archives from the previous release are reused: the packed mira conda environment, codec step-125000, LingBot/Wan runtime assets, and the complete 27,500-episode latent/text cache.

The raw RoboTwin2.0 dataset is intentionally not included. Download or mount the 100 ALOHA ZIP files separately at datasets/RoboTwin2.0; World v6 needs their original three-view RGB frames when building Qwen anchors.

Current restore set

Archive Restores Status
source-code-world-v6-20260805.tar.zst current mira/, lingbot-vision/, v6 plans new
mira-conda-env.tar.zst relocatable Python 3.12 / Torch 2.8 CUDA 12.8 environment reused
runtime-models.tar.zst LingBot-Vision Large and Wan UMT5 assets reused
qwen3-vl-2b-instruct.tar.zst pinned frozen Qwen3-VL-2B current encoder new
codec-checkpoint-step125000.tar.zst codec weights plus optimizer/training state reused
world-checkpoint-step80000.tar.zst complete 32-shard audited v5 World migration source new
world-derived-cache.tar.zst complete latent/text cache and generated metadata reused
world-v6-migration-inputs.tar.zst EMA selection, state encoder, smoke reports/videos new

Historical source, World-50k, and individually uploaded World-80k objects remain in repository history for compatibility, but the current restore script does not download or extract them.

All current payloads are covered by SHA256SUMS. Exact source/archive sizes, hashes, model contracts, and reused/new status are recorded in PACKAGE_MANIFEST.json. No credential is included.

Direct selective download

Install a recent Hugging Face CLI, clear every proxy, and download only the current restore set:

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 PACKAGE_MANIFEST.json SHA256SUMS 'skills/**' \
    archives/source-code-world-v6-20260805.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 ask an agent to use skills/multi-wam-migrate/SKILL.md. The deterministic restore entry point is:

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

After downloading/mounting the ALOHA data and creating a destination-local SwanLab token, run verify_install.sh and start_world_v6_smoke.sh. The smoke launcher uses only physical GPUs 6 and 7. Formal World v6 remains an explicitly guarded eight-GPU pipeline.

Official dataset: https://huggingface.co/datasets/TianxingChen/RoboTwin2.0
Official RoboTwin code: https://github.com/RoboTwin-Platform/RoboTwin

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