🤖 Quick-Bridge-Streamer: A BridgeData V2 RLDS Viewer

demo.mp4

"Preview BridgeData without polluting your PyTorch environment."

A lightweight BridgeData V2 viewer for PyTorch users who want to inspect TFRecord datasets without installing full TensorFlow or risking CUDA conflicts.

✨ Features

• ⚡ PyTorch-Safe — TensorFlow-CPU only (GPU disabled). Zero CUDA conflicts with your PyTorch environment.
• 🎥 Real-time visualization — High-performance streaming via Rerun SDK
• 🔍 Smart analytics — Detects frozen robots and gripper spam automatically
• 🎯 PyTorch code generation — Auto-generates ready-to-use Dataset code on exit
• 🌐 Web-based viewer — Opens in browser automatically, no CLI tools needed

🚀 Quick Start

Step 1: Clone & Install

git clone https://github.com/kwfang/Quick-Bridge-Streamer.git
cd Quick-Bridge-Streamer
pip install -r requirements.txt  # Safe: only tensorflow-cpu, no CUDA deps

Step 2: Prepare Your Data

Place your BridgeData V2 TFRecord files in a folder (e.g., data/):

data/
├── bridge_dataset-train.tfrecord-00000-of-XXXXX
├── bridge_dataset-train.tfrecord-00001-of-XXXXX
└── dataset_info.json

Step 3: Run

python main.py data --limit 100

Parameters:

• data: Path to dataset folder (containing .tfrecord-* files)
• --limit: Number of frames to stream (-1 = unlimited, default)
• --split: Data split (train or val, default: train)
• --no-interactive: Auto-select first dataset without prompting

Step 4: Exit

When you press Ctrl+C to exit, the terminal automatically prints:

[SUCCESS] Session ended.

Need to load this data in PyTorch? Here is your snippet:
---------------------------------------------------------
class BridgeDataV2Dataset(Dataset):
    def __init__(self, data_path="...", split="train"):
        ...
    def __getitem__(self, idx):
        ...
---------------------------------------------------------
Copy the above to start training!

📁 Project Structure

Quick-Bridge-Streamer/
├── oxview/                 # Core package
│   ├── __init__.py         # Version info
│   ├── loader.py           # TFRecord parsing & streaming
│   ├── viewer.py           # Rerun web visualization
│   ├── analytics.py        # Quality detection algorithms
│   └── exporter.py         # PyTorch code generation
├── main.py                 # CLI entry point
├── requirements.txt        # Dependencies
└── README.md               # This file

🔧 Advanced Usage

Scan multiple datasets

# If data/ contains multiple subfolders with datasets
python main.py data --no-interactive

Stream unlimited frames

python main.py data --limit -1

Select specific dataset

python main.py data --dataset 2 --limit 50

🧪 Dataset Format

BridgeData V2 structure:

Feature	Type	Shape	Description
action	float32	[7]	[pos_x, pos_y, pos_z, rot_x, rot_y, rot_z, gripper]
language_instruction	string	—	Natural language task description
observation/image_0	uint8	[256, 256, 3]	Main camera RGB (JPEG)
observation/state	float32	[7]	Robot joint angles + gripper position

🐛 Troubleshooting

ImportError: No module named 'tensorflow'

pip install -r requirements.txt

Browser doesn't open

Check firewall settings or manually open: http://localhost:9090

📄 License

MIT License — See LICENSE for details.

---

Version: 0.1.0