---
license: apache-2.0
task_categories:
  - robotics
  - depth-estimation
  - video-classification
tags:
  - stereo-depth
  - 4d-sound
  - manipulation
  - pick-and-place
  - kitchen
  - simulation
  - strands-robots
  - newton-physics
  - mujoco
  - nvidia
  - jetson-thor
  - blackwell
pretty_name: "Kitchen Stereo4D - Pick & Place with Spatial Audio"
size_categories:
  - 1K<n<10K
---

# 🍳 Kitchen Stereo4D Dataset

**Stereo depth + spatial audio pick-and-place dataset generated by [strands-robots](https://github.com/cagataycali/strands-gtc-nvidia) on NVIDIA Jetson AGX Thor**

## Overview

This dataset contains synchronized stereo RGB, metric depth maps, 3D point clouds, 
and binaural spatial audio from robot manipulation tasks in a kitchen environment.

| Metric | Value |
|--------|-------|
| Episodes | 10 |
| Total Frames | 501 |
| Resolution | 640×480 |
| Stereo Baseline | 50mm |
| FPS | 30 |
| Robot | Franka Emika Panda (simplified) |
| GPU | NVIDIA Thor (Blackwell, 132GB) |
| Simulator | MuJoCo 3.5.0 |

## Generation Pipeline

```
Kitchen Scene (MuJoCo) → Stereo Cameras → Depth Estimation → Point Cloud
         ↓                                        ↓                ↓
  Contact Events → Spatial Audio Synthesis    Disparity Maps    Dataset
         ↓                                        ↓                ↓
  Robot Trajectory → State Recording        Visualization   HuggingFace
```

## Structure

```
├── episode_000/
│   ├── stereo_left/        # Left camera RGB frames (PNG)
│   ├── stereo_right/       # Right camera RGB frames (PNG)  
│   ├── depth/              # Metric depth maps (NPY, float32, meters)
│   ├── disparity/          # Disparity maps (NPY, float32, pixels)
│   ├── point_clouds/       # 3D point clouds (NPY, float32, XYZ)
│   ├── audio/              # Binaural spatial audio (WAV, 44.1kHz)
│   ├── contacts.json       # Physics contact events
│   ├── robot_state.json    # Joint positions, velocities
│   └── metadata.json       # Episode metadata
├── ...
├── overview.png            # Scene overview render
├── side_view.png           # Side view render
├── metadata.json           # Dataset metadata
└── README.md               # This file
```

## Episodes

| Episode | Task | Frames | Contacts | Mode |
|---------|------|--------|----------|------|
| 0 | Pick up the mug and place it on the plate | 97 | 14918 | gpu_mujoco |
| 1 | Pick up the bowl and move it to the cutting board | 97 | 14810 | gpu_mujoco |
| 2 | Grasp the spatula and place it next to the bowl | 97 | 14954 | gpu_mujoco |
| 3 | Exploration trajectory 1 | 30 | 6253 | gpu_mujoco_variation |
| 4 | Exploration trajectory 2 | 30 | 6350 | gpu_mujoco_variation |
| 5 | Exploration trajectory 3 | 30 | 6406 | gpu_mujoco_variation |
| 6 | Exploration trajectory 4 | 30 | 6830 | gpu_mujoco_variation |
| 7 | Exploration trajectory 5 | 30 | 6786 | gpu_mujoco_variation |
| 8 | Exploration trajectory 6 | 30 | 6325 | gpu_mujoco_variation |
| 9 | Exploration trajectory 7 | 30 | 6325 | gpu_mujoco_variation |


## Spatial Audio

Audio is synthesized from physics contact events with binaural spatialization:
- **Ceramic** objects (mug, plate, bowl): High-frequency impacts (2200Hz base)
- **Wood** objects (cutting board, spatula): Mid-frequency (800Hz base)  
- **Metal** contacts (gripper): Sharp transients (4000Hz base)
- **Spatial panning**: Interaural Level Difference (ILD) based on contact position

## Usage

```python
import numpy as np
from PIL import Image

# Load a frame
left = np.array(Image.open("episode_000/stereo_left/frame_000000.png"))
right = np.array(Image.open("episode_000/stereo_right/frame_000000.png"))
depth = np.load("episode_000/depth/frame_000000.npy")
disparity = np.load("episode_000/disparity/frame_000000.npy")
point_cloud = np.load("episode_000/point_clouds/frame_000000.npy")

# Load audio
import soundfile as sf
audio, sr = sf.read("episode_000/audio/spatial_audio.wav")

# Load metadata
import json
with open("episode_000/metadata.json") as f:
    meta = json.load(f)
```

## Citation

```bibtex
@misc{strands_kitchen_stereo4d_2026,
    title={Kitchen Stereo4D: Pick-and-Place with Spatial Audio},
    author={strands-robots},
    year={2026},
    publisher={HuggingFace},
    url={https://huggingface.co/datasets/cagataydev/strands-kitchen-stereo4d},
    note={Generated on NVIDIA Jetson AGX Thor (Blackwell, 132GB)},
}
```

## Hardware

- **Device**: NVIDIA Jetson AGX Thor
- **GPU**: Blackwell (sm_110), 132GB unified memory
- **CUDA**: 13.0
- **PyTorch**: 2.12.0+cu130
- **MuJoCo**: 3.5.0

## Related

- [strands-robots](https://github.com/cagataycali/strands-gtc-nvidia) — Agentic robotics framework
- [nvidia/ffs_stereo4d](https://huggingface.co/datasets/nvidia/ffs_stereo4d) — Reference stereo4D format
- [nvidia/PhysicalAI-Robotics-Manipulation-Objects-Kitchen-MJCF](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Manipulation-Objects-Kitchen-MJCF)