Multi-WAM reproducible handoff

Latest code: 2026-11-07. release-20261107/ contains the updated training and inference pipeline, launchers, held-out mini-eval tools, and checksums. It is a code-only overlay on the 2026-11-05 release; the saved step-6563 checkpoint and its reported evaluation remain historical BGR artifacts.

Latest: 2026-11-05 backup of the final 4-GPU joint checkpoint and official full evaluation. release-20261105/ contains step 6563, its paired World/Action deployment, the older 8-GPU step 25000 and 30181 checkpoints, current code, split manifests, restore scripts, checksums, and official evaluation evidence including 10,000 episode videos. The final step-6563 evaluation with BGR-compatible input scored 72.09% over all 100 task/domain cases. See the release README for exact restore commands, external data requirements, and the documented historical training source limitation. Do not use the newer RGB training input mode for this BGR checkpoint's evaluation.

The 2026-10-18 eval bring-up documents the working single-GPU RoboTwin pipeline for the running joint step-25000 snapshot.

Historical 2026-10-14 release: the exact 8-rank step-17700 checkpoint (including optimizer, scheduler, sampler and RNG), source, run metadata and 8-GPU resume scripts are in release-20261014/. Training logged through step 17710 before stopping; step 17700 was the newest complete checkpoint in that release. The earlier release-20261010/ supplies the packed environments and reused upstream assets: World-v12 step 25066, Action-v13 step 23016, codec step 125000 and Qwen3-VL-2B-Instruct. Raw RoboTwin data is external and not packaged.

For the October 14 release, read its progress and reproduction guide, continuation skill, release manifest, and checksums. The October 10 progress report and its 24-rank release remain historical evidence.

This model trained on 50 RoboTwin 2.0 tasks, 26,950 train episodes (clean and randomized), three synchronized views, 20 Hz sampled images/actions, and 10 Hz codec latents. World+Action are trainable; Qwen and codec are frozen. The October 14 continuation used 8 GPUs, batch 15/GPU and a global batch of 120. The observed closed-loop result 4437/9600 = 46.22% on 96/100 completed task-domain groups was measured on the older step-12531 model; it is not a full paper-comparable C2C/C2R number.

Bootstrap the historical step-17700 release

Use direct Hub access, or on this cluster set HTTP_PROXY and HTTPS_PROXY to http://192.168.0.135:18000 before the first command. Clear any proxy pointing to the Codex-reserved 127.0.0.1:7892 port.

hf download Orangerl/multi-wam --revision snapshot-20261014-joint17700 \
  --include 'README.md' 'PROGRESS_20261014.md' \
  'release-20261014/scripts/**' --local-dir multi-wam-hf
cd multi-wam-hf
bash release-20261014/scripts/download_release.sh "$PWD"
bash release-20261014/scripts/restore_release.sh "$PWD" /path/to/empty/workspace
bash release-20261014/scripts/download_dataset.sh /path/to/empty/workspace
/path/to/empty/workspace/conda/envs/mira/bin/python \
  release-20261014/scripts/verify_release.py /path/to/empty/workspace --require-dataset

The bundled download helpers preserve only that permitted workspace proxy or use direct access; they never use 127.0.0.1:7892. No credentials are in the payload. Set destination-local SwanLab/HF credentials yourself. The 8-GPU preflight and resume commands are in the latest progress guide. The eval environment still needs a compatible NVIDIA driver and Vulkan renderer. Never treat the 96-group score as a complete official benchmark result.

The 2026-09-02 historical summary remains in PROGRESS_20260902.md; its release and other earlier snapshots are preserved.