Visualize Dataset (v2.0+ latest dataset format)

Visualize Dataset (v2.0+ latest dataset format)

Browse robotics datasets in the LeRobot format with synchronized video, interactive charts, 3D robot replay and tools to flag bad episodes. For robotics researchers and practitioners.

Made by lerobot on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: LeRobot Dataset Visualizer is a web app for exploring robotics datasets, especially those in the LeRobot format. You move between organizations, datasets and episodes, watch the video and read charts that move in step with it. It also gives statistics, ways to spot bad episodes and a 3D replay of the robot.

Who it suits: It suits people who work with robot recordings and want to look inside a dataset before they train on it.

Industry | Who uses it | How they use it | Example
Robotics | Researchers | Inspect large recorded datasets | Open an episode and watch video beside the charts
Machine learning | ML engineers | Judge data quality before training | Use the filtering panel to flag jerky episodes
Education | Students | Learn what robot data looks like | Step through a robot's joints in the 3D replay
Data labeling | Annotators | Edit language annotations by hand | Add a subtask note in the annotations panel

Skip it if your data is not robot episodes, since the tool is built around that kind of dataset.

Where to use it:

• A web browser, on the Space's page at Hugging Face
• Docker on your own Windows, macOS or Linux computer

Use case | Who | What they get back
Explore an episode | Researchers | Video playing in step with sensor and control charts
Clean up a dataset | ML engineers | A list of flagged episodes as a ready-to-run LeRobot command
Replay the robot in 3D | Roboticists | Joint poses frame by frame with an end-effector trail

How to set it up: The online page needs nothing. A local copy needs Docker, Python with pip to fetch the files, and the settings that the repository lists in .env.example.

• Docker
• Python and pip, to fetch the files
• The settings listed in .env.example
• Windows, macOS or Linux

1. Install the Hugging Face hub library.
2. Download the Space's files into a folder named visualize_dataset.
3. Move into that folder.
4. Build the Docker image.
5. Run the container with port 7860 published, then open that port in your browser.

pip install -U huggingface_hub
hf download lerobot/visualize_dataset --repo-type space --local-dir visualize_dataset
cd visualize_dataset
docker build -t visualize_dataset .
docker run -it -p 7860:7860 visualize_dataset

How to use it: Open the page, pick an organization, a dataset and an episode with the sidebar and the navigation controls. Watch the video, then open the panels you need.

docker run -it -p 7860:7860 visualize_dataset

Run that line for a local copy and browse to port 7860 on your machine.

Tool | What it does
Overview panel | Summarizes the dataset's metadata, camera info and episode details
Statistics panel | Reports how many episodes there are, the total time recorded, the frame rate and how episode lengths are spread
Action insights panel | Offers autocorrelation, state-action alignment, speed distribution and a variance heatmap to guide training choices
Filtering panel | Flags episodes with little movement, jerky motion or odd length, then exports their IDs as a LeRobot command
3D replay | Steps through joint poses one frame at a time for a set of supported robots, such as the SO-100 and the Unitree G1
3D hands | Finds native hand annotations and shows wrists, 21-joint hand skeletons and wrist trails
Annotations panel | Lets you edit language annotations by hand and draw boxes or keypoints on a camera view

Example: You open an episode and watch the camera video while the charts move with it. You then use the filtering panel to flag episodes with jerky motion, and copy the command it gives you for removing them.

For newcomers and experts: Start with the overview and statistics panels of a dataset you know. Then open one episode and watch how the video and the charts line up.

The app uses Next.js with React, Recharts for the charts and Three.js with react-three-fiber for the 3D scene. It loads parquet and JSON data with pagination, chunking and lazy panels, so large datasets open quickly. The hands view needs specific annotation fields, including 156-dimension actions and hand joints in camera and world coordinates. No model weights or meshes are required. A camera overlay is switched off when the camera, its calibration or the video size is missing or mismatched.

Pick an organization, dataset and episode
Metadata, video and telemetry load from the Hub
Panels draw charts, statistics and 3D views
You flag bad episodes or edit annotations

Good to know:

• Optional service: A FastAPI service in the backend folder rewrites parquet files and pushes to the Hub for annotation edits.
• Limits: The hands view needs the annotation fields listed above. Missing or masked poses stay as gaps.
• License: apache-2, with a LICENSE file in the repository.
• Cost to run: none named.
• Data it reaches: Dataset metadata, episode video and sensor data from the Hub. Annotation edits can be written back to the Hub through the optional service.

On Hugging Face, 13 November 2026

• Likes: 754
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/lerobot/visualize_dataset
• Files: https://huggingface.co/spaces/lerobot/visualize_dataset/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.