Document verified ALTER hardware v1 and pinned public code

NOTICE

@@ -9,4 +9,4 @@ Paper, website media, logos and third-party packages are outside this artifact
 license. The models use a torchvision ResNet18 architecture; the released source
 constructs it with weights=None. No separate pretrained DINO weights are included.
 Simulator software and robot/scene assets are installed separately, not bundled
-as standalone software/assets here. Hardware artifacts are not included.
+as standalone software/assets here. Hardware v1 artifacts are documented separately under hardware/v1/.

README.md

@@ -19,8 +19,8 @@ remain retained separately, and exports record original and export hashes.
 
 Use only with the matching model/preprocessing, camera and action conventions
 in the release source. Data are trusted Python pickle artifacts, not a safe
-format for arbitrary untrusted downloads. Hardware recordings and their trial
-evidence are not included. Original ALTER demonstrations, replay and accompanying
+format for arbitrary untrusted downloads. Hardware recordings are released separately under `hardware/v1/`; scored physical-trial
+evidence is not included. Original ALTER demonstrations, replay and accompanying
 original records are released under CC BY 4.0; see LICENSE and NOTICE. Third-party
 software, assets and the paper retain their separate terms.
 
@@ -32,3 +32,10 @@ retraining or full evaluation rerun is claimed by preparation tests.
 See USAGE.md for downloads and the accompanying code release status.
 
 Paper: [Residual Denoising Enables Sample-Efficient Multi-Agent Coordination on Demand](https://arxiv.org/abs/2609.32129).
+
+## Hardware v1
+
+Selected hardware models, source data and exact prepared caches are now available
+under [hardware/v1](hardware/v1/README.md). See the [hardware download guide](hardware/v1/USAGE.md)
+for the separate pinned manifest, selective downloads and offline validation.
+Simulation payloads and their pinned revisions remain unchanged.

USAGE.md

@@ -36,4 +36,9 @@ Materialization requires all four bundles. It verifies hashes, relocates portabl
 artifact references and strictly validates derived checkpoint contracts. Direct
 use of portable statistics before materialization is not supported. Only load
 trusted pickle/PyTorch files. Short tests check functionality; they do not rerun
-full training or establish paper success rates. Hardware is deferred.
+full training or establish paper success rates. Hardware v1 is documented separately in [hardware/v1/USAGE.md](hardware/v1/USAGE.md).
+
+## Hardware release code
+
+Use [hardware v1 instructions](hardware/v1/USAGE.md) and the [tested code revision](https://github.com/labicon/ALTER/tree/a4919df35d703fafcf054717f8a53fd876b6a010).
+Hardware and simulation use separate download directories and manifests.

hardware/v1/README.md

@@ -0,0 +1,37 @@
+# ALTER hardware data
+
+Two independently selectable data bundles are included:
+
+| Bundle | Files | Bytes | Contents |
+| --- | ---: | ---: | --- |
+| hardware-demonstrations-data | 501 | 48,137,006,286 | Selected base demonstrations, paired demonstrations, and distilled replay |
+| hardware-training-data | 412 | 20,798,158,144 | Original `images.npy` / `trajectory.npz` for 206 prepared records |
+
+The model/metadata bundle provides the portable manifests and catalogs. There are
+332 base-training and 32 base-validation records; the high-data adaptation set
+contains 69 paired demonstrations and 68 replay records. The low-data set uses
+30 pairs and 30 replay records from the same cache pool. Shared data are stored
+once. Cardboard replay recordings were policy-generated, not human demonstrations.
+
+High-data sampling has 93,164 two-arm anchors and 44,224 replay anchors; low-data
+sampling has 37,319 and 19,921. Original timestamps, indices, phase labels, weights,
+action values, image bytes, normalization metadata, grouping and ordering are
+preserved. Cartesian xyz uses millimeters; rotations use radians; action rows are
+`[x, y, z, roll, pitch, yaw, gripper]`.
+
+These are recovered prepared inputs, not a complete raw hardware backup. Some
+original raw recordings and scored physical-trial evidence are absent. Some
+prepared records use retained-row ordinal time rather than physical timestamps;
+their original per-record provenance documents this. The current-frame training
+protocol is preserved rather than reinterpreted as a temporal dataset.
+
+Original data/cache bytes are unchanged; export receipts and checksum manifests
+record their identities. Path-bearing metadata is sanitized in independent copies.
+Representative inspection of 59 frames showed robot workspaces and task objects,
+with no obvious faces or private text in those samples. This was a representative review, not an exhaustive inspection of every frame.
+
+Published as hardware v1. See [USAGE.md](USAGE.md) and [release-manifest.json](release-manifest.json)
+for downloads pinned to immutable artifact revisions. Existing repository license
+files are unchanged by this addition. The simulation artifacts remain unchanged.
+
+Authors: Dayi Dong, Maulik Bhatt, Aayushi Shrivastava, Lasse Peters, Negar Mehr.

hardware/v1/USAGE.md

@@ -0,0 +1,44 @@
+# ALTER hardware v1 downloads
+
+Code: https://github.com/labicon/ALTER/tree/a4919df35d703fafcf054717f8a53fd876b6a010
+Models: https://huggingface.co/Berkeley-ICON-Lab/ALTER-models/tree/main/hardware/v1
+Data: https://huggingface.co/datasets/Berkeley-ICON-Lab/ALTER-data/tree/main/hardware/v1
+
+Use the hardware release branch and its Python 3.10 offline installation profile:
+
+```bash
+git clone https://github.com/labicon/ALTER.git
+cd ALTER
+git checkout a4919df35d703fafcf054717f8a53fd876b6a010
+```
+
+From that checkout, download and validate the model-only bundle (about 0.82 GB):
+
+```bash
+python scripts/download_release.py --manifest release/hardware-manifest.json \
+  --bundles hardware-models --destination /path/to/ALTER-hardware-artifacts
+python scripts/materialize_hardware_release.py \
+  --bundle-root /path/to/ALTER-hardware-artifacts \
+  --output "$PWD/public-validation/hardware-models" --bundles hardware-models
+python scripts/validate_hardware_models.py \
+  --root "$PWD/public-validation/hardware-models/hardware" \
+  --output "$PWD/public-validation/hardware-model-check.json"
+```
+
+For adaptation training, add `hardware-training-data` to the downloader and
+materializer bundle lists (about 20.80 GB additional data). For original selected
+base demonstrations and replay recordings, add `hardware-demonstrations-data`
+(about 48.14 GB). Use a fresh materialization output directory for each selection.
+Use a separate hardware download directory from the simulation download tree;
+the two releases have distinct root control manifests and receipts.
+
+The high/low training manifests are at `hardware/manifests/high.json` and
+`hardware/manifests/low.json` in the materialized output. See the public hardware
+guide for short CPU training commands, explicit base checkpoint/statistics
+arguments and NumPy pickle compatibility. Missing historical provenance remains
+explicit; no full paper reproduction or physical robot validation is claimed.
+
+The pinned manifest seals payload checksums and export receipts. Downloads reuse
+valid files and reject conflicting files. Only load trusted pickle/PyTorch
+artifacts. Seven original models and six short training workflows passed offline
+checks. Physical operation requires the separate private ICON_Arm project.