File size: 3,557 Bytes
71022b5
 
3319dfc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71022b5
3319dfc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
---
license: mit
task_categories:
- robotics
tags:
- robotics
- imitation-learning
- in-context-learning
- lerobot
- franka
- manipulation
- human-demonstrations
language:
- en
pretty_name: SimpleICL Real-Robot Data Sample
size_categories:
- n<1k
---

# SimpleICL Real-Robot Data Sample

A sample of our paired human-robot manipulation data, released alongside
[SimpleICL](https://github.com/simpleicl/simple-icl) — the official code release of
the paper *In-context Robot Learning Made Simple: A Democratized Recipe for
Manipulation Tasks*. It is a subset of the full dataset collected for
the project, provided to illustrate the data format and to accompany the
training recipe in the repository README (*Training*).

## Dataset Summary

100 paired episodes of bimanual Franka FR3 tabletop manipulation. In every
pair, a teleoperated robot arm and a human hand perform the same task back to
back; the human-side demonstration video serves as the in-context condition
(`video_ctx`) that the policy consumes at training and inference time.

| | |
| --- | --- |
| Episodes | 100 pairs (100 robot + 100 human) |
| Total frames (robot side) | 28,457 @ 30 fps |
| Episode length | 157–449 frames |
| Robot side | 3 RealSense cameras (high + left/right wrist), 848×480 |
| Human side | 1 high camera, 848×480 |
| Action / state | 16-D (7 joints + 1 gripper per arm) |

Scene alterations keep the task unchanged and perturb only task-irrelevant
factors — object positions, lighting, and background colors. The same task
recorded under the base and altered placements references each other as
in-context conditions (see `robot/meta/conditions_list.json`), so a policy
trained on this data learns *what the task is* from the demonstration video
while staying robust to where objects sit, how the scene is lit, and what the
background looks like.

## Structure

The data is distributed as a single archive, `pair_data.zip`. Unzipping it
produces two sibling LeRobot-format (v2.1) dataset directories:

```
pair_data.zip
└── (extracts to:)
    ├── robot/                      # teleoperated robot episodes
    │   ├── meta/                   # tasks, episodes, scenes, conditions
    │   │   ├── tasks.jsonl
    │   │   ├── episodes.jsonl
    │   │   ├── episodes_stats.jsonl
    │   │   ├── info.json
    │   │   ├── scenes.json          # episode index -> scene placement label
    │   │   ├── plan_rows.json
    │   │   ├── conditions.json
    │   │   └── conditions_list.json # episode index -> human demo videos (video_ctx)
    │   ├── data/                   # parquet per episode
    │   └── videos/                 # 3 cameras per episode
    └── human/                      # human-hand demonstrations
        └── videos/                 # cam_high video per episode
```

Condition paths in `robot/meta/conditions_list.json` are relative to the
extraction root, so the dataset keeps working after the extracted directory
is moved as a whole.

## Usage

Download the archive, unzip it, and point the SimpleICL training config at
the robot side — see
the repository [SimpleICL](https://github.com/simpleicl/simple-icl) README for the one-command training flow.

## Additional Information

- **License**: MIT
- **Origin**: a sample subset of the data collected by the SimpleICL authors
  on a dual-arm Franka FR3 platform with RealSense cameras.
- **Associated code**: [SimpleICL](https://github.com/simpleicl/simple-icl)