y0urOy/FastWAM-PiperX

FastWAM — RoboDojo Piper X

Fixed step_028630 checkpoint for one multitask Piper X policy covering all six RoboDojo real-robot tasks. Source and installation · Paired policy

Native FastWAM video-denoising baseline trained on the same Piper X episodes, normalization, initialization and update budget as JAM.

Files

• assets/weights/fastwam/step_028630.pt — original checkpoint, 12,041,813,369 bytes, SHA256 383863b3c0f6d387e6a6cafcb34e2a65e52f11ebaf637d4ddf71b25cc9a86511.
• models/fastwam/deployment.json — fixed inference configuration.
• models/fastwam/dataset_stats.json — original training-only normalization.
• provenance.json, FILE_MANIFEST.json, SHA256SUMS — source, training and content checksums.

The shared ActionDiT initialization is stored once in JAM-Wan-PiperX. Download it to the same workspace before loading this policy. Public Wan base assets are downloaded from pinned upstream revisions; they are not duplicated here.

Download and run

Clone the code, install the runtime using its Piper X instructions, then merge this snapshot into JAM/integrations/robodojo:

from huggingface_hub import snapshot_download, hf_hub_download
workspace = "/path/to/JAM/integrations/robodojo"
snapshot_download("y0urOy/FastWAM-PiperX", local_dir=workspace,
                  allow_patterns=["assets/weights/fastwam/*", "models/fastwam/*"])
hf_hub_download("y0urOy/JAM-Wan-PiperX",
                "assets/base/ActionDiT_linear_interp_Wan22_alphascale_1024hdim.pt", local_dir=workspace)

For reproducible transfer, set revision to this release's Hub commit and verify SHA256SUMS. All three binary files retain their original content hashes.

cd /path/to/JAM/integrations/robodojo
POLICY_DIR="$PWD/XPolicyLab/policy/JAM"
python3.10 "$POLICY_DIR/setup_workspace.py"
# Download the pinned Wan2.2/UMT5/VAE and Wan2.1 tokenizer assets:
python3.10 "$POLICY_DIR/assets.py" download-public
python3.10 "$POLICY_DIR/assets.py" verify --method fastwam
# Runtime must already be installed as described in the code repository:
bash "$POLICY_DIR/serve.sh" fastwam cap_pen /opt/venvs/piperx-policy 9010 127.0.0.1 0

The service binds to loopback. Real-robot access uses the official platform protocol once confirmed. This is a PyTorch policy checkpoint loaded by the supplied adapter, not a Transformers AutoModel.

Tasks and fixed settings

cap_pen, classify_objects, disassemble_LEGO, hang_mugs, pack_objects_into_backpack, sweep_blocks.

Three RGB cameras plus current 14D joint state and instruction; bf16, seed 42, 10 inference steps, action horizon 32, replan 24. Outputs are absolute joint targets: left arm (6), left gripper (1), right arm (6), right gripper (1). The data were recorded at 25 Hz; controller cadence and unit/gripper conventions await platform confirmation.

Training: 540 train / 60 held-out episodes, per-task 90/10 split, seed 42, global batch 128 on eight GH200 nodes, 10 epochs / 28,630 updates. Dataset revision: 35efbc7dedfdbeeb6e95fb749bd885d73d483e41.

Evaluation status

Real-robot registration was received on 2026-10-16. Piper X six-task acceptance, protocol and schedule remain pending. Real-robot trials: 0. Success rate: null / pending. The 240 offline action chunks per model are interface/offline checks and are not robot trials. This release makes no claim of official leaderboard results or measured task success.

The original candidate passed historical Miyabi raw-image, encoded-image and batch interface checks. Fresh platform installation, end-to-end robot trials and task success are unverified. See the code's validation report.

License

Model files are provided under Apache-2.0 with retained Wan attribution in NOTICE; JAM source additions are MIT. The public Wan models and RoboDojo dataset retain their own terms. No raw training dataset, training-state snapshots or applicant contact records are included.