Multi-WAM reproducible handoff

Latest: 2026-10-10, joint World+Action step 12531. The exact 24-rank training checkpoint (including optimizer, scheduler, sampler and RNG), its deploy export, source, two packed conda environments, split/statistics, eval code/evidence and videos are in release-20261010/. The earlier snapshot-20260902-full remains immutable and supplies four 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.

Read progress, the reproduction skill, release manifest, and checksums. The old September training skill has been updated to route to the new protocol; old release files are historical, not the latest objective.

This model trained on 50 RoboTwin 2.0 tasks, 26,950 train episodes (clean and randomized), three synchronized views, 20 Hz sampled RGB/action, and 10 Hz codec latents. World+Action are trainable; Qwen and codec are frozen. The run migrated per-GPU batch 3→5 at step 2449 with optimizer state retained, then stopped at step 12531 of a configured 55000 steps. The observed closed-loop result is 4437/9600 = 46.22% on 96/100 completed task-domain groups, not a full paper-comparable C2C/C2R number; see progress for the four simulator failures and protocol caveats.

Bootstrap on another machine (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-20261010-joint12531 \
  --include 'README.md' --include 'PROGRESS_20261010.md' \
  --include 'skills/multi-wam-joint-reproduce/**' \
  --local-dir multi-wam-hf
cd multi-wam-hf
bash skills/multi-wam-joint-reproduce/scripts/download_release.sh "$PWD"
bash skills/multi-wam-joint-reproduce/scripts/restore_release.sh "$PWD" /path/to/empty/workspace
bash skills/multi-wam-joint-reproduce/scripts/verify_release.sh /path/to/workspace
bash skills/multi-wam-joint-reproduce/scripts/download_dataset.sh /path/to/workspace

No credentials are in the payload. Set destination-local SwanLab/HF credentials yourself. 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.