Datasets:

Modalities:
Tabular
Formats:
json
Size:
< 1K
ArXiv:
License:
File size: 2,282 Bytes
dc19166
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
# 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.