Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,21 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
## Dataset Description
|
2 |
+
|
3 |
+
StaticEmbodiedBench is a dataset for evaluating vision-language models on embodied intelligence tasks, as featured in the OpenCompass leaderboard
|
4 |
+
(https://staging.opencompass.org.cn/embodied-intelligence/rank/brain).
|
5 |
+
|
6 |
+
It covers three key capabilities:
|
7 |
+
|
8 |
+
- **Macro Planning**: Decomposing a complex task into a sequence of simpler subtasks.
|
9 |
+
- **Micro Perception**: Performing concrete simple tasks such as spatial understanding and fine-grained perception.
|
10 |
+
- **Stage-wise Reasoning**: Deciding the next action based on the agent’s current state and perceptual inputs.
|
11 |
+
|
12 |
+
Each sample is also labeled with a visual perspective:
|
13 |
+
|
14 |
+
- **First-Person View**: The visual sensor is integrated with the agent, e.g., mounted on the end-effector.
|
15 |
+
- **Third-Person View**: The visual sensor is separate from the agent, e.g., top-down or observer view.
|
16 |
+
|
17 |
+
This release includes **200 open-source samples** from the full dataset, provided for public research and benchmarking purposes.
|
18 |
+
|
19 |
+
---
|
20 |
+
license: mit
|
21 |
+
---
|