accebet commited on
Commit
7ae6dc9
Β·
verified Β·
1 Parent(s): b408d42

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +23 -43
README.md CHANGED
@@ -1,31 +1,3 @@
1
- ---
2
- configs:
3
- - config_name: specs
4
- default: true
5
- data_files:
6
- - split: full
7
- path: viewer_data/specs.parquet
8
- - config_name: checklists
9
- data_files:
10
- - split: full
11
- path: viewer_data/checklists.parquet
12
- - config_name: qa_items
13
- data_files:
14
- - split: full
15
- path: viewer_data/qa_items.parquet
16
- - config_name: video_sources
17
- data_files:
18
- - split: full
19
- path: viewer_data/video_sources.parquet
20
- - config_name: source_annotations
21
- data_files:
22
- - split: full
23
- path: viewer_data/source_annotations.parquet
24
- - config_name: rules
25
- data_files:
26
- - split: full
27
- path: viewer_data/rules.parquet
28
- ---
29
  <!-- markdownlint-disable MD013 MD033 MD041 MD060 -->
30
 
31
  <p align="center">
@@ -36,7 +8,7 @@ configs:
36
 
37
  <div align="center">
38
 
39
- 🌐 [Homepage](https://csu-jpg.github.io/LabInstruct.github.io/) | πŸ‘‰ [Dataset](#data-files) | πŸ“„ Paper (coming soon) | πŸ’» [Code](https://github.com/CSU-JPG/LabInstruct) | πŸ† [Leaderboard](#leaderboard)
40
 
41
  </div>
42
 
@@ -73,8 +45,12 @@ configs:
73
 
74
  ## πŸ“œ Abstract
75
 
 
 
76
  Self-driving laboratories (SDLs) aim to automate the full experimental loop, from scientific decision-making to physical execution. Ideally, AI-generated plans could be carried out directly by robotic systems, but reliable automation remains difficult in complex, open-world laboratory environments, where experiments often involve fine-grained manipulation, long-horizon procedures, and substantial variation across tasks and setups. Humans therefore remain an important execution interface between AI-generated plans and physical experiments, creating a need for clear and effective human-facing experimental guidance. Because laboratory procedures are inherently visual, spatial, and dynamic, video is particularly well suited to communicating apparatus configurations, manipulation actions, temporal dependencies, and state changes. Recent advances in video generation now make it possible to synthesize experimental demonstrations directly from an initial workspace image and a natural-language instruction. However, whether such models can reliably communicate real laboratory procedures has not been systematically studied. We introduce **LabInstruct**, a benchmark for situated instructional video generation in real laboratories. LabInstruct contains 204 tasks across 5 scientific disciplines, with real reference executions and structured annotations of objects, actions, contacts, and state transitions. Evaluating 8 frontier image-to-video models, we find that visually plausible generations frequently remain procedurally incorrect, revealing a substantial gap between visual realism and the reliability required for experimental instruction.
77
 
 
 
78
  <a id="project-overview"></a>
79
 
80
  ## 🌟 Project Overview
@@ -228,18 +204,21 @@ These prompts are written to produce structured drafts for human verification.
228
 
229
  ## πŸ† Leaderboard
230
 
231
- Pooled Overall is evaluated by GPT-5.6 Sol at 4 FPS. Higher is better; † marks commercial models.
 
 
 
 
 
 
 
 
 
 
 
232
 
233
- | Rank | Model | Type | Organization | Pooled Overall |
234
- | :--: | :-- | :-- | :-- | --: |
235
- | **1** | **MiniMax H3** | Open weight | MiniMax | **46.3** |
236
- | 2 | Seedance 2.0 † | Commercial | ByteDance | 45.1 |
237
- | 3 | Wan 3.0 † | Commercial | Alibaba | 44.4 |
238
- | 4 | Wan 2.2-I2V-A14B | Open weight | Alibaba | 25.9 |
239
- | 5 | Cosmos3 Super | Open weight | NVIDIA | 22.8 |
240
- | 6 | LTX 2.3 | Open weight | Lightricks | 22.1 |
241
- | 7 | LingBot Video | Open weight | Robbyant | 19.5 |
242
- | 8 | Cosmos3 Nano | Open weight | NVIDIA | 19.4 |
243
 
244
  <a id="bibtex"></a>
245
 
@@ -252,10 +231,11 @@ If you find our work helpful, please consider citing it:
252
  title = {LabInstruct: Benchmarking Situated Instructional Video Generation for Lab Procedures},
253
  author = {Yuming Fu and Weijia Wu and Jing Chen and Jiahao Tang and Feifei Chen and Hongyu Zhu and Xin Jin and Alex Jinpeng Wang},
254
  year = {2026},
255
- eprint = {XXXX.XXXXX},
256
  archivePrefix = {arXiv},
257
- primaryClass = {cs.CV},
258
- url = {https://arxiv.org/abs/XXXX.XXXXX}
 
259
  ```
260
 
261
  <a id="contact"></a>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  <!-- markdownlint-disable MD013 MD033 MD041 MD060 -->
2
 
3
  <p align="center">
 
8
 
9
  <div align="center">
10
 
11
+ 🌐 [Homepage](https://csu-jpg.github.io/LabInstruct.github.io/) | πŸ‘‰ [Dataset](#data-files) | πŸ“„ [Paper](https://arxiv.org/abs/2610.14020) | πŸ’» [Code](https://github.com/CSU-JPG/LabInstruct) | πŸ† [Leaderboard](#leaderboard)
12
 
13
  </div>
14
 
 
45
 
46
  ## πŸ“œ Abstract
47
 
48
+ <div align="justify">
49
+
50
  Self-driving laboratories (SDLs) aim to automate the full experimental loop, from scientific decision-making to physical execution. Ideally, AI-generated plans could be carried out directly by robotic systems, but reliable automation remains difficult in complex, open-world laboratory environments, where experiments often involve fine-grained manipulation, long-horizon procedures, and substantial variation across tasks and setups. Humans therefore remain an important execution interface between AI-generated plans and physical experiments, creating a need for clear and effective human-facing experimental guidance. Because laboratory procedures are inherently visual, spatial, and dynamic, video is particularly well suited to communicating apparatus configurations, manipulation actions, temporal dependencies, and state changes. Recent advances in video generation now make it possible to synthesize experimental demonstrations directly from an initial workspace image and a natural-language instruction. However, whether such models can reliably communicate real laboratory procedures has not been systematically studied. We introduce **LabInstruct**, a benchmark for situated instructional video generation in real laboratories. LabInstruct contains 204 tasks across 5 scientific disciplines, with real reference executions and structured annotations of objects, actions, contacts, and state transitions. Evaluating 8 frontier image-to-video models, we find that visually plausible generations frequently remain procedurally incorrect, revealing a substantial gap between visual realism and the reliability required for experimental instruction.
51
 
52
+ </div>
53
+
54
  <a id="project-overview"></a>
55
 
56
  ## 🌟 Project Overview
 
204
 
205
  ## πŸ† Leaderboard
206
 
207
+ GPT-5.6 Sol is evaluated at 4 FPS; Human is the pooled overall human rating. Higher is better; † marks commercial models.
208
+
209
+ | Rank | Model | Type | Organization | GPT-5.6 Sol | Human | Tier |
210
+ | :--: | :-- | :-- | :-- | --: | --: | :--: |
211
+ | **1** | **MiniMax H3** | Open weight | MiniMax | **46.3** | **44.3** | **T0** |
212
+ | 2 | Seedance 2.0 † | Commercial | ByteDance | 45.1 | 43.0 | T0 |
213
+ | 3 | Wan 3.0 † | Commercial | Alibaba | 44.4 | 37.3 | T0 |
214
+ | 4 | Wan 2.2-I2V-A14B | Open weight | Alibaba | 25.9 | 15.4 | T1 |
215
+ | 5 | Cosmos3 Super | Open weight | NVIDIA | 22.8 | 11.5 | T1 |
216
+ | 6 | LTX 2.3 | Open weight | Lightricks | 22.1 | 3.0 | T2 |
217
+ | 7 | LingBot Video | Open weight | Robbyant | 19.5 | 10.5 | T2 |
218
+ | 8 | Cosmos3 Nano | Open weight | NVIDIA | 19.4 | 5.8 | T2 |
219
 
220
+ > [!NOTE]
221
+ > Model rankings shift slightly across evaluation views, evaluators, sampling rates, and generation seeds, so we also report a Tier to keep the comparison robust.
 
 
 
 
 
 
 
 
222
 
223
  <a id="bibtex"></a>
224
 
 
231
  title = {LabInstruct: Benchmarking Situated Instructional Video Generation for Lab Procedures},
232
  author = {Yuming Fu and Weijia Wu and Jing Chen and Jiahao Tang and Feifei Chen and Hongyu Zhu and Xin Jin and Alex Jinpeng Wang},
233
  year = {2026},
234
+ eprint = {2610.14020},
235
  archivePrefix = {arXiv},
236
+ primaryClass = {cs.RO},
237
+ url = {https://arxiv.org/abs/2610.14020}
238
+ }
239
  ```
240
 
241
  <a id="contact"></a>