# Physical AI: Unreal Engine - Isaac Sim Navigation Dataset

Demonstration datasets for the Unitree G1 humanoid robot performing vision-based object navigation in procedurally generated indoor environments.
Physics calculation is performed using Isaac Sim while the rendering is performed by Unreal Engine.
We assume the dataset is placed inside `demo_data` in [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T).

## Task

The robot receives a front-facing camera image and a language instruction specifying a target object in the scene. It must navigate to the target object using a combination of high-level velocity commands (HLC) and low-level joint position actions (LLC).

Example instruction:
```
Scene contains: Plant in a Pot, Towel, Aerosol Spray Can, Light Blue Bucket, Old Clock. Navigate to the Old Clock.
```

## Dataset Variants

| Dataset | Episodes | Description |
|---|---|---|
| `g1_procedural_room_navigation_20260206_062009` | 100 | 5 objects per scene |
| `g1_procedural_room_navigation_20260206_080307` | 100 | 1 object per scene |
| `g1_procedural_room_navigation_20260206_095145` | 100 | 3 objects per scene |

## Dataset Format

Each dataset follows the [LeRobot v2](https://github.com/huggingface/lerobot) format:

```
g1_procedural_room_navigation_*/
├── meta/
│   ├── info.json          # Schema, features, robot config, processing params
│   ├── episodes.jsonl     # Per-episode metadata (index, length, task instruction)
│   ├── tasks.jsonl        # Task index definitions
│   ├── modality.json      # Modality-to-column mapping with slice indices
│   └── stats.json         # Per-feature statistics (see Generating Statistics section below)
├── data/
│   └── chunk-{NNN}/
│       └── {episode_index:06d}.parquet
└── videos/
    └── chunk-{NNN}/
        └── observation.images.front/
            └── episode_{episode_index:06d}.mp4
```

## Features

| Feature | Type | Shape | Description |
|---|---|---|---|
| `observation.images.front` | video | (480, 640, 3) | Front camera RGB at 50 fps |
| `observation.state.joint_pos` | float32 | (29,) | Joint positions (rad) |
| `observation.state.joint_vel` | float32 | (29,) | Joint velocities (rad/s) |
| `observation.state.root_pos_w` | float32 | (3,) | Root position in world frame |
| `observation.state.root_quat_w` | float32 | (4,) | Root orientation quaternion (w, x, y, z) |
| `observation.state.root_lin_vel_b` | float32 | (3,) | Root linear velocity in body frame |
| `observation.state.root_ang_vel_b` | float32 | (3,) | Root angular velocity in body frame |
| `action.hlc_raw` | float32 | (3,) | Raw high-level command (vx, vy, omega_z) |
| `action.hlc_processed` | float32 | (3,) | Processed HLC (scaled, shifted, clipped) |
| `action.llc_raw` | float32 | (29,) | Raw low-level joint position targets |
| `action.llc_processed` | float32 | (29,) | Processed LLC (scaled around default pose) |
| `timestamp` | float64 | (1,) | Time in seconds from episode start |
| `episode_id` | int64 | (1,) | Episode index |
| `frame_id` | int64 | (1,) | Frame index within episode |

## Robot

- **Model**: Unitree G1
- **Joints**: 29 DoF (legs, waist, arms, wrists)
- **Joint order**: IsaacLab convention
- **FPS**: 50

## Combining Datasets

To merge multiple collection sessions into a single dataset, edit `SOURCE_DATASETS` and `OUTPUT_DATASET` in the script, then run:

```bash
python demo_data/scripts/combine_datasets.py
```

This will:
- Re-index episodes continuously (0, 1, ..., N-1) across all sources
- Copy parquet files with updated `episode_id` columns
- Symlink video files to the originals (no duplication)
- Merge `episodes.jsonl` with new indices
- Create `meta/origin.yaml` tracking which source datasets were combined
- Correctly bucket episodes into `chunk-NNN/` directories when total episodes exceed `chunks_size`

## Generating Statistics

After combining (or for any new dataset), generate `stats.json` using:

```bash
python gr00t/data/stats.py <dataset_path> --embodiment-tag <embodiment-tag>
```

For example:

```bash
python gr00t/data/stats.py demo_data/g1_procedural_room_navigation_combined --embodiment-tag unitree_g1_navigation_vel
```

This computes per-feature statistics (mean, std, min, max, q01, q99) across all parquet files and writes them to `meta/stats.json`. It also generates `meta/relative_stats.json` for relative action representations if configured in the embodiment config.

**Note**: The embodiment's modality configuration must be defined in `gr00t/configs/data/embodiment_configs.py` and the tag must be added in `gr00t/data/embodiment_tags.py` before running this script.