---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- GR00T
- GR00T-N1.7
- robotics
- imitation-learning
---

# Robocasa — LeRobot/GR00T format

Robocasa converted to the **GR00T-N1.7** training format: **LeRobot v2.1** directory
layout plus `meta/modality.json`. Validated by a 5-step GR00T-N1.7 fine-tuning smoke-train
(loss decreased, checkpoint written) after conversion.

## Contents
- **Episodes:** 20
- **Frames:** 7,372
- **codebase_version:** v2.1 (+ `meta/modality.json` for GR00T)
- **Features:** 3 cameras (agentview left/right + eye-in-hand), 16-dim state, 12-dim action

> Note: this is a **validation subset** used to verify the v3→GR00T conversion pipeline
> end-to-end, not the full source dataset.

## Source
Converted from **DAVIAN-Robotics/robocasa-H50** (LeRobot v3.0) using the
[lerobot-groot-dataset-conversion](https://github.com/nv-sachdevkartik/lerobot-groot-dataset-conversion)
pipeline, which wraps NVIDIA Isaac-GR00T's conversion scripts.

## Usage
Load with GR00T-N1.7's `LeRobotSingleDataset` using `embodiment_tag=new_embodiment` and the matching modality config.