|
|
--- |
|
|
license: mit |
|
|
task_categories: |
|
|
- robotics |
|
|
- reinforcement-learning |
|
|
tags: |
|
|
- act |
|
|
- imitation-learning |
|
|
- mujoco |
|
|
- simulation |
|
|
- aloha |
|
|
size_categories: |
|
|
- n<1K |
|
|
--- |
|
|
|
|
|
# 🦾 ACT Simulation Dataset (v2): Cube Sort |
|
|
|
|
|
**A lightweight, randomized imitation learning dataset optimized for Action Chunking Transformers.** |
|
|
|
|
|
<div align="center"> |
|
|
<video src="https://github-production-user-asset-6210df.s3.amazonaws.com/71585678/520477863-ec1d9885-0d81-46a1-a29d-c0ef0e782fff.mp4?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20251201%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20251201T004617Z&X-Amz-Expires=300&X-Amz-Signature=dd18d7f406bcc24fbd9ddde59686ea3984949a407477cf45acfd2e2f9a709370&X-Amz-SignedHeaders=host" width="400" controls></video> |
|
|
</div> |
|
|
|
|
|
|
|
|
|
|
|
## 📖 Overview |
|
|
This dataset contains **228 episodes** of synthetic robot manipulation data generated using MuJoCo. It was created to train **Action Chunking Transformers (ACT)** on consumer hardware (specifically an Apple M2 MacBook Air). |
|
|
|
|
|
Unlike "sterile" datasets that follow perfect straight lines, this **v2 dataset** utilizes **Domain Randomization** to ensure the policy learns robust correction behaviors rather than simple memorization. |
|
|
|
|
|
- **Task:** A 4-DOF Robotic Arm picking up a cube and moving it to a specific target zone. |
|
|
- **Hardware Virtualization:** Mimics the data structure of a real Dynamixel-based robot arm. |
|
|
- **Goal:** Enable "Zero-to-Hero" training of Visuomotor Policies without NVIDIA H100s. |
|
|
|
|
|
## ⚙️ Dataset Specifications |
|
|
|
|
|
| Feature | Details | |
|
|
| :--- | :--- | |
|
|
| **Episodes** | 228 (approx. 68,000 timesteps) | |
|
|
| **Format** | `.hdf5` | |
|
|
| **Simulation** | MuJoCo (via `dm_control` / `mujoco-py`) | |
|
|
| **Robot** | Custom 4-DOF Arm (Base, Shoulder, Elbow, Wrist) | |
|
|
| **Camera** | 1x Static Front Camera (480x640, RGB) | |
|
|
| **Control Freq** | 50Hz | |
|
|
|
|
|
### 🧠 Data Structure (HDF5) |
|
|
Each `.hdf5` file represents one full episode and contains: |
|
|
|
|
|
- `observations/images/front`: `(T, 480, 640, 3)` - Raw RGB images. |
|
|
- `observations/qpos`: `(T, 4)` - Joint positions (radians). |
|
|
- `observations/qvel`: `(T, 4)` - Joint velocities. |
|
|
- `action`: `(T, 4)` - Target joint positions for the next step. |
|
|
|
|
|
## 🛠️ How to Use |
|
|
|
|
|
### 1. Download & Unzip |
|
|
This dataset is stored as a compressed ZIP file (`episodes-v2.zip`) to maintain directory structure. |
|
|
|
|
|
```python |
|
|
from huggingface_hub import hf_hub_download |
|
|
import zipfile |
|
|
import os |
|
|
|
|
|
# 1. Download |
|
|
zip_path = hf_hub_download( |
|
|
repo_id="sanskxr02/act_sim_cube_sort", |
|
|
filename="episodes-v2.zip", |
|
|
repo_type="dataset" |
|
|
) |
|
|
|
|
|
# 2. Extract |
|
|
extract_path = "data/sim_cube_sort" |
|
|
os.makedirs(extract_path, exist_ok=True) |
|
|
|
|
|
with zipfile.ZipFile(zip_path, 'r') as zip_ref: |
|
|
zip_ref.extractall(extract_path) |
|
|
|
|
|
print(f"✅ Dataset ready at {extract_path}") |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
