README.md

---
language:
- en
tags:
- robotics
- soft-robotics
- sim2real
- physics-simulation
- neural-networks
- pneumatic-actuation
- motion-capture
- surrogate-modeling
- fem-simulation
- sofa-framework
size_categories:
- 100K<n<1M
task_categories:
- tabular-regression
- time-series-forecasting
task_ids:
- tabular-single-column-regression
- univariate-time-series-forecasting
pretty_name: "Soft Manipulator Sim2Real Dataset"
configs:
- config_name: default
  data_files:
  - "*.csv"
dataset_info:
  features:
  - name: P1
    dtype: float64
    description: "Pneumatic pressure for cavity 1 (Pa)"
  - name: P2
    dtype: float64
    description: "Pneumatic pressure for cavity 2 (Pa)"  
  - name: P3
    dtype: float64
    description: "Pneumatic pressure for cavity 3 (Pa)"
  - name: thetaX
    dtype: float64
    description: "Joint angle X-axis (radians)"
  - name: thetaY
    dtype: float64
    description: "Joint angle Y-axis (radians)"
  - name: d
    dtype: float64
    description: "Linear displacement (mm)"
  - name: TCP_X
    dtype: float64
    description: "Tool center point X position (mm)"
  - name: TCP_Y
    dtype: float64
    description: "Tool center point Y position (mm)"
  - name: TCP_Z
    dtype: float64
    description: "Tool center point Z position (mm)"
  splits:
  - name: train
    num_bytes: 167000000
    num_examples: 200000
  download_size: 167000000
  dataset_size: 167000000
license: mit
paperswithcode_id: null
---


# SoftManipulator Sim2Real Dataset

This dataset accompanies the research paper "Bridging High-Fidelity Simulations and Physics-Based Learning Using A Surrogate Model for Soft Robot Control" published in Advanced Intelligent Systems, 2025.

## 📋 Dataset Overview

This dataset contains experimental and simulation data for a 3-actuator pneumatic soft manipulator, designed to enable sim-to-real transfer learning and surrogate model development. The data includes motion capture recordings, pressure mappings, SOFA FEM simulation outputs, and surrogate model training datasets.

## 🎯 Dataset Purpose

- **Sim2Real Research**: Bridge the gap between SOFA simulations and real hardware
- **Surrogate Model Training**: Train neural networks for fast dynamics prediction
- **Model Calibration**: Calibrate FEM parameters using real-world data
- **Workspace Analysis**: Understand the robot's range of motion and capabilities
- **Validation**: Compare simulation outputs with experimental ground truth

## 📊 Dataset Files

| File | Size | Samples | Description | Usage |
|------|------|---------|-------------|--------|
| `ForwardDynamics_Pybullet_joint_to_pos.csv` | ~66MB | 100,000+ | PyBullet forward dynamics: joint commands → TCP positions | Surrogate model training |
| `MotionCaptureData_ROM.csv` | ~15MB | 10,000+ | Real robot motion capture trajectories | Ground truth validation |
| `PressureThetaMappingData.csv` | ~2MB | 5,000+ | Pressure inputs → joint angle outputs | Actuation mapping |
| `Pressure_vs_TCP.csv` | ~8MB | 8,000+ | Pressure commands → tool center point positions | Control modeling |
| `RealPressure_vs_SOFAPressure.csv` | ~3MB | 3,000+ | Hardware vs simulation pressure comparison | Model calibration |
| `SOFA_snapshot_data.csv` | ~45MB | 50,000+ | FEM nodal displacements from SOFA simulations | Physics validation |
| `SurrogateModel_ROM.csv` | ~12MB | 15,000+ | Reduced-order model training data | Fast inference |
| `SurrogateModel_withTooltip_ROM.csv` | ~18MB | 20,000+ | ROM data with tooltip contact forces | Contact modeling |

## 🔧 Data Collection Setup

### Hardware Configuration
- **Robot**: 3-cavity pneumatic soft manipulator (silicone, ~150mm length)
- **Actuation**: Pneumatic pressure control (-20 kPa to +35 kPa per cavity)
- **Sensing**: 6-DOF motion capture system (OptiTrack), pressure sensors
- **Materials**: Ecoflex 00-30 silicone with embedded pneumatic chambers

### Simulation Environment
- **SOFA Framework**: v22.12 with SoftRobots plugin
- **FEM Model**: TetrahedronFEMForceField with NeoHookean material
- **Material Properties**: Young's modulus 3-6 kPa, Poisson ratio 0.41
- **PyBullet**: v3.2.5 for surrogate model validation

## 📈 Data Schema

### Joint Space Data
- `thetaX`, `thetaY`: Joint angles (radians, -π/4 to π/4)
- `d`: Linear displacement (mm, 0 to 50)

### Pressure Commands
- `P1`, `P2`, `P3`: Cavity pressures (Pa, -20000 to 35000)

### Cartesian Space
- `TCP_X`, `TCP_Y`, `TCP_Z`: Tool center point position (mm)
- `Normal_X`, `Normal_Y`, `Normal_Z`: End-effector orientation

### Forces
- `Fx`, `Fy`, `Fz`: External forces (N, contact/manipulation tasks)

### Temporal Information
- `Time`: Timestamp (seconds)
- `Episode`: Experiment episode number

## 🚀 Usage Examples

### Loading Data in Python
```python
import pandas as pd
from datasets import load_dataset

# Load from HuggingFace
dataset = load_dataset("Ndolphin/SoftManipulator_sim2real")

# Or load locally
df = pd.read_csv("ForwardDynamics_Pybullet_joint_to_pos.csv")
print(f"Dataset shape: {df.shape}")
print(f"Columns: {df.columns.tolist()}")
```

### Training a Surrogate Model
```python
# Pressure to joint angle mapping
X = df[['P1', 'P2', 'P3']].values  # Pressure inputs
y = df[['thetaX', 'thetaY', 'd']].values  # Joint outputs

from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

# Train your neural network model
```

### Motion Analysis
```python
# Analyze workspace coverage
import matplotlib.pyplot as plt

tcp_data = df[['TCP_X', 'TCP_Y', 'TCP_Z']]
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.scatter(tcp_data['TCP_X'], tcp_data['TCP_Y'], tcp_data['TCP_Z'])
ax.set_title('Robot Workspace')
```

## 📝 Data Quality & Preprocessing

### Quality Assurance
- **Filtering**: Outliers removed using 3-sigma rule
- **Smoothing**: Savitzky-Golay filter applied to motion capture data
- **Synchronization**: All sensors synchronized to 100Hz sampling rate
- **Validation**: Cross-validated against multiple experimental runs

### Recommended Preprocessing
```python
from sklearn.preprocessing import StandardScaler

# Normalize features for neural network training
scaler = StandardScaler()
X_normalized = scaler.fit_transform(X)

# Save scaler for inference
import joblib
joblib.dump(scaler, 'scaler.pkl')
```

## 🎓 Citation

If you use this dataset in your research, please cite:

```bibtex
@article{hong2025bridging,
  title={Bridging High-Fidelity Simulations and Physics-Based Learning Using A Surrogate Model for Soft Robot Control},
  author={Hong, T. and Lee, J. and Song, B.-H. and Park, Y.-L.},
  journal={Advanced Intelligent Systems},
  year={2025},
  publisher={Wiley}
}
```

## 📄 License

This dataset is released under the MIT License. See LICENSE file for details.

## 🤝 Contact

For questions about the dataset or research:
- **Authors**: T. Hong, J. Lee, B.-H. Song, Y.-L. Park
- **Institution**: [Your Institution]
- **Email**: [Contact Email]
- **Paper**: [ArXiv/DOI Link when available]

## 🔍 Related Resources

- **Code Repository**: https://github.com/ndolphin-github/Sim2Real_framework_SoftRobot
- **SOFA Simulations**: Included in the repository
- **Pre-trained Models**: Available in the code repository
- **Demo Videos**: SOFA simulation demos included

---

*Dataset Version: 1.0 | Last Updated: October 2025*