episodes list | seed int64 | train_episode_indexes list | validation_episode_indexes list |
|---|---|---|---|
[
{
"archive": "arm08/episodes/episode_000000.tar",
"arm_id": "arm08",
"frames": 1260,
"invalid_images": 35,
"merged_episode_index": 0,
"source_episode_index": 0,
"source_session": "arm08_peg_insert_leader02_25_20260922_193503",
"split": "validation"
},
{
"archive": "arm08/epis... | 42 | [
1,
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17,
19,
20,
21,
22,
23,
25,
26,
29,
30,
31,
32,
34,
35,
36,
37,
38,
39,
40,
41,
42,
43,
44,
46,
47,
48,
49,
50,
51,
52,
53,
54,
57,
58,
59,
60,
61,
62,
63,
... | [
0,
2,
18,
24,
27,
28,
33,
45,
55,
56,
83,
95
] |
SO-101 peg insertion demonstrations
104 operator-kept episodes, 55,770 synchronized control frames, collected on four SO-101 follower arms with leader02 on September 22-23, 2026 (America/Los_Angeles).
Task: pick up the peg from the table, insert it fully into the fixed socket, release it, and retreat. Recording uses manual S/E/R episode boundaries; scene-reset motion is excluded. All saved episodes are included. Success was not independently visually certified.
| Arm | Episodes | Frames | Invalid images |
|---|---|---|---|
| arm08 | 26 | 15605 | 452 |
| arm09 | 25 | 14693 | 724 |
| arm10 | 28 | 13807 | 899 |
| arm11 | 25 | 11665 | 20 |
Layout and extraction
Each arm has its own folder. armXX/episodes/episode_NNNNNN.tar contains the original parquet, front/wrist JPEG frames and per-episode SHA-256 manifest, with original relative paths. armXX/meta/ contains the dataset and episode metadata. Extract all episode archives into their corresponding arm folder to recover the native recording layout. SHA256SUMS verifies the archives. This is the native fleet recording format with deferred JPEG images, not a stock LeRobot video export.
Each parquet contains calibrated LeRobot-normalized six-joint observations and absolute target actions, wall-clock timestamps, frame indexes, policy version and terminal fields. Camera dimensions and joint names are specified by each arm's meta/info.json. No time compaction or image interpolation is applied.
Image quality and training split
Original damaged images are retained for provenance and explicitly listed in invalid_images.json and each arm's quality.json. Never silently decode them with a tolerant decoder. The local DP preparation marks invalid observation frames; both training and validation exclude any sample whose observation history touches one. Actions and the original time sequence remain intact.
episode_catalog.csv and training_split.json map original arm-local episode indexes to combined indexes. The seeded split has 92 training and 12 validation episodes, with every arm represented in both. Normalization is fitted on training episodes only.
The local training recipe uses front/wrist RGB plus joint state, two observations, sixteen predicted absolute joint targets and eight executed actions. All model and optimizer settings are recorded with the training run. Offline validation loss does not establish physical insertion success.
- Downloads last month
- 683