Create dataset card with metadata and links to paper/code
Browse filesThis PR initializes the dataset card for EBench. It links the repository to the official paper, the project website, and the GitHub repository, and sets the appropriate task category and license in the YAML metadata.
README.md
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
task_categories:
|
| 4 |
+
- robotics
|
| 5 |
+
tags:
|
| 6 |
+
- lerobot
|
| 7 |
+
- vla
|
| 8 |
+
- mobile-manipulation
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# EBench: Elemental Mobile Manipulation Benchmark
|
| 12 |
+
|
| 13 |
+
This repository contains the training trajectories (stored in LeRobot format) for **EBench**, an indoor VLA (Vision-Language-Action) manipulation benchmark built on NVIDIA Isaac Sim. Instead of compressing a model's behavior into a single overall success rate, EBench produces a multi-axis capability profile that exposes what a model is good at and where it overfits.
|
| 14 |
+
|
| 15 |
+
- **Paper:** [EBench: Elemental Diagnosis of Generalist Mobile Manipulation Policies](https://huggingface.co/papers/2606.18239)
|
| 16 |
+
- **Project Page:** [EBench Homepage](https://internrobotics.github.io/EBench-home/)
|
| 17 |
+
- **Documentation:** [EBench Docs](https://internrobotics.github.io/EBench-doc/)
|
| 18 |
+
- **GitHub Repository:** [InternRobotics/EBench](https://github.com/InternRobotics/EBench)
|
| 19 |
+
|
| 20 |
+
## Key Features
|
| 21 |
+
|
| 22 |
+
- **Three manipulation regimes in one benchmark** — covers *long-horizon*, *dexterous & precise*, and *mobile* manipulation.
|
| 23 |
+
- **5-axis atomic diagnostic** — every task is labelled by *Scene · Atomic Skill · Horizon · Precision · Mobility*.
|
| 24 |
+
- **4-axis generalization tests** — controlled perturbations along *Object · Background · Instruction · Mixed*.
|
| 25 |
+
- **LeRobot Format** — trajectories are structured using Hugging Face's `lerobot` data format, making them ready for loading and training.
|
| 26 |
+
|
| 27 |
+
## Citation
|
| 28 |
+
|
| 29 |
+
If you find EBench useful in your research, please cite:
|
| 30 |
+
|
| 31 |
+
```bibtex
|
| 32 |
+
@misc{ebench2026,
|
| 33 |
+
title = {EBench: Elemental Mobile Manipulation Benchmark},
|
| 34 |
+
author = {Shanghai AI Laboratory},
|
| 35 |
+
year = {2026},
|
| 36 |
+
note = {Preprint coming soon},
|
| 37 |
+
url = {https://internrobotics.github.io/EBench-doc/}
|
| 38 |
+
}
|
| 39 |
+
```
|