File size: 2,764 Bytes
fccf3f1
 
9148e52
 
 
fccf3f1
9148e52
 
 
 
 
 
8ba021a
 
10d2cc7
9148e52
 
 
10d2cc7
9148e52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
tags:
- robotics
- in-context-learning
---
# Semantic-Discrimination-Test-Set

This repository releases the test resources of the paper [“In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks”](https://arxiv.org/abs/2609.38173)

## Semantic Discrimination Test Set

We do not construct our dataset using the full set of RoboTwin tasks, as fragile objects such as `bottle` are particularly prone to collisions in the dense-object settings designed for ICL, which can interfere with task execution.

The test data is collected via our tailored SimpleICL data pipeline, which is not the official RoboTwin setting.

### Directory structure

Our SimpleICL data pipeline produces episodes that are organized as:

```text
<scene_name>/
  placement0/			# One scene -> Multi-placement
    <task0_name>/		# One placement -> Multi-task
      <embodiment>[__pert0]/	# One task -> Multi-perturbation
        data/
          episode<I>.hdf5		
        instructions/
          episode<I>.json
        video/
          episode<I>.mp4		# One perturbation -> Multi-visual-variant
        scene_info.json	# Task-level meta info
      <embodiment>[__pert1]/
      ...
    <task1_name>/
    ...
    placement_meta.json	# Placement-level meta info
  placement1/
  ...
  completion.json 		# Collection completion markup
  manifest.json 		# Scene configuration
```

In this test set, each test episode is reorganized as:

```text
<task_name>__<serial_number>/
  placement0/
    <task_name>/<embodiment>
        data/
        instructions/
        video/
        scene_info.json
      placement_meta.json
  completion.json
  manifest.json
```

This removes all tasks unrelated to the evaluation, saves storage space, and still allows the framework designed for the aforementioned data structures to read the data.

## Project

- Project page: https://simpleicl.github.io/simpleicl/
- Paper: https://arxiv.org/abs/2609.38173
- Code repository: https://github.com/simpleicl/simple-icl

## Citation

```bibtex
@article{li2026context,
  title={In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks},
  author={Li, Minxing and Han, Minghao and Zhao, Weizhi and Wang, Hanwen and Liu, Xiangshuo and Shang, Shuyao and Zhou, Jingxiang and Sun, Mingchao and Pan, Hongyu and Xu, Mu and others},
  journal={arXiv preprint arXiv:2609.38173},
  year={2026}
}
```

## License

This release is a derivative of RoboTwin 2.0 framework for ICL-oriented data collection.

It is released under the MIT license, consistent with the upstream RoboTwin 2.0 release. Please also refer to the original RoboTwin project and dataset for upstream attribution and terms.

> RoboTwin repository: https://github.com/RoboTwin-Platform/RoboTwin