kanishqgandharv/olsd-v0.1

---
license: apache-2.0
task_categories:
- reinforcement-learning
tags:
- robotics
- locomotion
- mujoco
- gymnasium
- benchmark
- offline-rl
language:
- en
size_categories:
- 1M<n<10M
pretty_name: Open Locomotion Skills Dataset (OLSD)
---

# Open Locomotion Skills Dataset (OLSD) v0.1

A unified benchmark dataset for legged robot locomotion across diverse morphologies.

## Dataset Description

OLSD provides expert, random, and domain-randomized trajectories for training and evaluating
locomotion policies. All data is collected from MuJoCo/Gymnasium environments.

| Metric | Value |
|--------|-------|
| Total Episodes | ~4,200 |
| Total Steps | ~2,722,712 |
| Robots | HalfCheetah, Ant, Walker2d, Hopper |
| Quality Tiers | random, expert, domain_random |
| Format | Parquet + metadata JSON |

## Benchmark Results

| Robot | Algorithm | Norm. Score | Return |
|-------|-----------|-------------|--------|
| Hopper | PPO | **108.4** | 3,507.9 |
| Walker2d | SAC | **107.3** | 4,928.5 |
| Ant | PPO | 29.7 | 1,733.2 |
| HalfCheetah | PPO | 18.0 | 1,955.4 |

## Usage

```python
from datasets import load_dataset

# Load the full dataset
ds = load_dataset("kanishqgandharv/olsd-v0.1")

# Or load a specific robot
ds = load_dataset("kanishqgandharv/olsd-v0.1", data_files="halfcheetah/**/*.parquet")
```

Or with the OLSD SDK:

```python
import olsd
dataset = olsd.load("./data/olsd-v0.1-final")
print(dataset.summary())
```

## Data Schema

Each row contains:
- `observation.joint_positions` -- Joint angles (rad)
- `observation.joint_velocities` -- Joint velocities (rad/s)
- `action` -- Motor commands
- `reward` -- Reward signal
- `done` -- Terminal flag

## Links

- [GitHub](https://github.com/KanishqGandharv219/open-locomotion-dataset)
- [Website](https://kanishqgandharv219.github.io/open-locomotion-dataset/)
- [Quickstart Notebook](https://github.com/KanishqGandharv219/open-locomotion-dataset/blob/main/notebooks/quickstart.ipynb)

## License

Apache 2.0