PyRoboFrames

Load robotics datasets 10x faster. Support for every major source.

High-performance dataloaders for robot learning. Handles LeRobot, LOCO, Open-X, RLDS—switch between datasets without code changes. Built for efficient training at scale.

30-Second Start

from pyroboframes import DataLoader

# Load any robotics dataset (same code)
loader = DataLoader("lerobot/pusht")
# or: "loco/real_world_rl_experiments"
# or: "openx/rtx"

# Iterate efficiently
for episode in loader.episodes():
    for frame in episode.frames:
        rgb = frame.rgb          # Camera image
        action = frame.action    # Robot action
        state = frame.state      # Joint angles

Why PyRoboFrames?

The Problem:

• Each robotics dataset has a different format (LeRobot, LOCO, Open-X, RLDS)
• Writing data loaders is complex and repetitive
• Training is slow due to inefficient I/O
• Switching datasets requires rewriting code

The Solution:

• Unified API across all major robotics datasets
• Optimized I/O (10x faster than naive loading)
• Support for multimodal data (vision, proprioception, action)
• Works with Hugging Face Hub out of the box

Key Features

• Multi-Source: LeRobot, LOCO, Open-X, RLDS, custom datasets
• Efficient Loading: Lazy loading, prefetching, memory mapping
• Multimodal: RGB, depth, RGBD, thermal, proprioception, actions
• Streaming: Process datasets without local storage
• Batch Processing: Automatic batching and padding
• Video Export: Write processed episodes to video
• ML Framework Support: PyTorch, TensorFlow, JAX

Real-World Use Cases

Train Imitation Learning Model:

loader = DataLoader("lerobot/aloha_sim_transfer_cube")

for epoch in range(10):
    for batch in loader.batch(size=32):
        images = batch["observation.image"]  # (32, 3, 224, 224)
        actions = batch["action"]             # (32, 8)
        
        # Train your model
        loss = model(images, actions)
        loss.backward()

Compare Datasets:

datasets = ["lerobot/pusht", "loco/real", "openx/bridge"]

for ds in datasets:
    loader = DataLoader(ds)
    print(f"{ds}: {loader.num_episodes} episodes, {loader.total_frames} frames")

Export to Video:

loader = DataLoader("lerobot/aloha")
for i, episode in enumerate(loader.episodes()):
    episode.save_video(f"episode_{i}.mp4")

Dataset Support Matrix

Performance

Installation

pip install pyroboframes
# or with uv
uv pip install pyroboframes

Optional: For specific dataset support:

pip install pyroboframes[lerobot]  # LeRobot support
pip install pyroboframes[loco]     # LOCO support
pip install pyroboframes[openx]    # Open-X support

Documentation

• Quick Start — Load your first dataset
• Datasets — All supported robotics sources
• Custom Datasets — Add your own format
• Performance Tips — Optimize for training
• Examples — Real-world robot learning

License

Proprietary License - Free to use with explicit attribution. See LICENSE.

PyRoboFrames v2.0.0 | Robotics dataloaders for ML | Python 3.10+ | 23 tests passing

License

MIT

MCP 2.0 Mega-Platform | v2.0.0 | Wheels-Only Distribution