kushsaraf/robot-intelligence-dataset

---
license: cc-by-4.0
task_categories:
- tabular-classification
- reinforcement-learning
language:
- en
tags:
- robot
- sensors
- activity-recognition
- navigation
- failure-detection
- reinforcement-learning
- tabular
pretty_name: Robot Intelligence Dataset
size_categories:
- 10K<n<100K
---

# 🤖 Robot Intelligence Dataset

A collection of three real-world robotics sensor datasets used to train and evaluate machine learning pipelines across four intelligence challenges:
**Perception · Navigation · Failure Detection · Autonomous Decision-Making (RL)**.

Companion GitHub repo: [KushSaraf/Robot_Intelligence](https://github.com/KushSaraf/Robot_Intelligence)

> **Made by:** Kush Saraf & Yash Chavda

---

## Datasets

### 🧠 1. Perception Dataset — Human Activity Recognition (HAR)
**Source:** [UCI HAR Dataset](https://archive.ics.uci.edu/ml/datasets/human+activity+recognition+using+smartphones)

| Property | Value |
|---|---|
| Instances | 10,299 (7,352 train / 2,947 test) |
| Features | 561 time & frequency domain features |
| Sensors | Accelerometer + Gyroscope (3-axis, 50 Hz) |
| Classes | 6 (Walking, Walking Upstairs, Walking Downstairs, Sitting, Standing, Laying) |
| Device | Samsung Galaxy S II worn on waist |
| Subjects | 30 volunteers, ages 19–48 |

**Task:** Classify human activity from smartphone inertial sensor signals.

**Files:**
```
perception_dataset/
├── train/X_train.txt # 7352 × 561 feature matrix
├── train/y_train.txt # labels 1–6
├── train/subject_train.txt # subject IDs
├── test/X_test.txt
├── test/y_test.txt
├── test/subject_test.txt
├── features.txt # feature names
├── activity_labels.txt # class names
└── train/Inertial Signals/ # raw inertial signals (9 channels)
```

---

### 🗺️ 2. Navigation Dataset — Wall-Following Robot
**Source:** [UCI Wall-Following Robot Navigation](https://archive.ics.uci.edu/ml/datasets/Wall-Following+Robot+Navigation+Data)

| Property | Value |
|---|---|
| Instances | 5,456 |
| Features | 24 ultrasonic range sensors (US1–US24, 360°) |
| Classes | 4 movement decisions |
| Robot | SCITOS G5 |
| Sampling | ~9 samples/second |

**Class distribution:**
| Class | Count | % |
|---|---|---|
| Move-Forward | ~2,205 | 40% |
| Sharp-Right-Turn | ~2,074 | 38% |
| Slight-Right-Turn | ~818 | 15% |
| Slight-Left-Turn | ~359 | 6% |

**Task:** Sensor array → movement command (non-linearly separable by design).

**Files:**
```
navigation_dataset/
├── sensor_readings_24.data # 24 sensors (recommended)
├── sensor_readings_4.data # 4 simplified sensors
├── sensor_readings_2.data # 2 simplified sensors
└── Wall-following.names # dataset description
```

---

### ⚠️ 3. Failure Detection Dataset — Robot Execution Failures
**Source:** [UCI Robot Execution Failures](https://archive.ics.uci.edu/ml/datasets/Robot+Execution+Failures)

| Property | Value |
|---|---|
| Instances | ~463 total across 5 tasks |
| Raw features | 90 (15 time-steps × 6 force/torque axes) |
| Stat features | 24 (mean, std, min, max per axis) |
| Total features | 114 (after feature engineering) |
| Classes | Up to 11 per file |

**Force/torque axes:** Fx, Fy, Fz, Tx, Ty, Tz

**Learning problems (files):**
| File | Task |
|---|---|
| lp1.data | Approach to grasp position |
| lp2.data | Transfer of a part |
| lp3.data | Position after transfer failure |
| lp4.data | Approach to ungrasp position |
| lp5.data | Motion with a part |

**Task:** Classify robot arm execution failures from time-series force/torque data.

**Files:**
```
failure_dataset/
├── lp1.data # approach to grasp
├── lp2.data # transfer
├── lp3.data # post-transfer
├── lp4.data # ungrasp approach
└── lp5.data # motion with part
```

---

## Usage

### Perception
```python
import pandas as pd
X_train = pd.read_csv("perception_dataset/train/X_train.txt", sep=r"\s+", header=None)
y_train = pd.read_csv("perception_dataset/train/y_train.txt", header=None, names=["label"])
```

### Navigation
```python
import pandas as pd
df = pd.read_csv("navigation_dataset/sensor_readings_24.data", header=None)
X, y = df.iloc[:, :-1], df.iloc[:, -1]
```

### Failure Detection
```python
# Blocks are separated by blank lines; each block = one instance
# Format: label\n6_values\n6_values\n...\n (15 timesteps)
def parse_lp(path):
rows = []
with open(path) as f:
text = f.read()
for block in text.strip().split("\n\n"):
lines = block.strip().split("\n")
label = lines[0].strip()
data = [list(map(float, l.split())) for l in lines[1:] if l.strip()]
rows.append({"label": label, "data": data})
return rows
```

---

## Citation

If you use this dataset collection, please cite the original UCI sources:

- HAR: Anguita et al., ESANN 2013
- Wall-Following: Freire et al., ESANN 2009
- Robot Failures: Camarinha-Matos et al., 1996