🍬 RoboCandyWrapper

Sweet wrappers for extending and remixing LeRobot Datasets.

---

🍬 Why do I need this?

You have robot data. Lots of it. But working with it is a pain.

Your datasets are split across incompatible LeRobot versions, extending or transforming them risks breaking compatibility, and balancing across data sources takes more effort than it should.

RoboCandyWrapper handles all of this:

• Mix datasets freely — Load v2.1 and v3.0 LeRobot datasets through a single unified interface, and use them together as if they were the same format.

• Extend without breaking — Add custom labels or columns to existing datasets via Plugins, while staying fully compatible with LeRobot tooling.

• Control your data mix — Use built-in Samplers to increase or decrease the weight of specific datasets in your mix.

⚠️ RoboCandyWrapper is still experimental so do note that the API could change in the future, although we'll do our best to avoid unnecessary changes!

🍬 Quick Start (5 Minutes)

Installation

# Include LeRobot as a dependency in installation
pip install robocandywrapper

# OR...
# Use your own version of LeRobot - may cause issues!
pip install --no-dependencies robocandywrapper

# OR...
# Use your own version of LeRobot and install robocandywrapper as a local editable dependency so you change LeRobot imports as needed
# This might be required if you use a LeRobot fork or depend on an out of date version
git clone https://github.com/villekuosmanen/RoboCandyWrapper.git
cd RoboCandyWrapper
pip install --no-dependencies -e .

Basic usage

Load a vintage v2.1 dataset and a modern v3.0 dataset as if they were the same thing.

from robocandywrapper import make_dataset_without_config

# Your playlist: one old, one new
repo_ids = [
    "lerobot/svla_so100_pickplace",  # v2.1 dataset
    "lerobot/svla_so100_stacking",   # v3.0 dataset
]

# The factory handles the compatibility logic automatically
dataset = make_dataset_without_config(repo_ids)

print(f"🎉 Successfully loaded {len(dataset)} episodes from mixed sources!")

🍬 What more can I do with it?

🎧 The "Mix Tape" (Mixing Datasets)

Learn how to combine multiple datasets into one, handle different robot configurations, and use sampling weights to balance your data mix.

🧂 The "Flavor Enhancer" (Transforming Data)

Learn how to use Plugins to add new labels or columns to your dataset, reshape tensors, or modify existing data on-the-fly without breaking backwards compatability.

Other cool stuff from the authors

1. Physical AI Interpretability offers open-source interpretability tools for AI robotics.
2. RewACT is an open-source reward model / value function based on the ACT transformer architecture.