01
Workflows
Programmable processing
Standardization, quality checks, and post-processing are each a replaceable algorithm. Use the built-in ones, or write your own in Python or as a container image, and arrange them in the visual editor.
Programmable Data Platform for Embodied AI
EmbodiFlow
From recordings to trained weights. Robot teleoperation, human egocentric, and UMI recordings are processed, versioned, and trained on one platform, and every step can run your own algorithm.
Overview
EmbodiFlow takes over once recording is done. Recordings from TeleXperience and SenseXperience, as well as ROS bag, MCAP, HDF5, LeRobot, and UMI data, can be uploaded and processed by programmable workflows into dataset versions that are ready to train on and reproducible.
See It In Action
EmbodiFlow in Action
Follow the main path from upload and programmable processing to dataset versions and model training.
Platform
Process, version, and train on one platform
Uploaded recordings run through a workflow, are frozen into a dataset version, and are trained on to produce downloadable weights.
02
Versions
Frozen, reproducible datasets
A version never changes once frozen, and new data never alters an old version. Export LeRobot v3, RLDS, Zarr, or HDF5, with checksummed files.
03
Training
From version to trained weights
Train ACT, SmolVLA, OpenPI, and other policies on a dataset version, watch training as it runs, and download the weights when it finishes.
Core Features
One platform from recording to robot
Your workflow, your algorithms
Every step is a replaceable algorithm: use ours, or bring your own. Python or container images, a visual workflow editor, and GPUs, retries, and caching handled by the platform.
Robots, humans, and UMI in one format
Data from every source becomes LeRobot v3 on one 30 Hz grid, ready to train together. Includes actions from egocentric video, UMI calibration and SLAM, depth, hand pose, and face anonymization.
Bad data stays out
Failing episodes are dropped, not whole recordings, and anything unchecked never shows as passed. Every finding comes with evidence and links straight to the moment in replay.
Every model can say what made it
Versions are frozen and every run is pinned and reproducible. Every model traces back to the data, algorithms, and parameters that produced it.
AI does the data work
AI cuts long recordings into episodes and proposes their task text, and you confirm. Every AI proposal stays marked until it is reviewed.
From version to trained weights
Train ACT, SmolVLA, or OpenPI on a version, then download the weights. Or export LeRobot v3, RLDS, Zarr, or HDF5, with one download per version.
Open and Deployable
Works with your tools, runs where you need it
Bring your AI assistant
AI coding tools and other assistants work the platform through MCP, the CLI, and SDKs: upload data, write checks, start training, and compare results.
Works with the formats you record
ROS bag, MCAP, HDF5, LeRobot, UMI / action camera, egocentric headsets, RLDS, and Zarr.
Three ways to run it
SaaS Team runs in the cloud with nothing to operate. Workstation is a GPU workstation with the platform preinstalled, fully offline. Enterprise runs on your private cloud for multiple teams, with high availability and custom features.
Documentation & Guides
Documentation
- Start
Quickstart
Upload your first recording, review its episodes and quality, then freeze a version and export it.
- Guide
Build your own workflow
Combine built-in and custom algorithms in the visual editor.
- Guide
Write a custom algorithm
Write standardization, check, and post-processing algorithms in Python or as a container image.
- Guide
Quality checks
Review each finding and its evidence, and jump to the matching moment in replay.
- Guide
Train a model
Train a policy on a frozen version, watch it learn, and download the weights.
Deployment sizing
Turn recordings into data you can train on. Size a SaaS, workstation, or private-cloud deployment for your data.
Frequently asked questions
- Which teams is EmbodiFlow for?
- Robot learning teams in university labs, robotics companies, and corporate research groups, as well as data collection and data service teams. Use it to organize your own recordings, run your own processing algorithms, train and compare policies, and deliver datasets by version.
- Which data formats are supported?
- You can upload ROS bag, MCAP, HDF5, LeRobot, UMI / action camera, egocentric headset, RLDS, and Zarr data, which is unified as LeRobot v3 on a 30 Hz grid. Exports include LeRobot v3, RLDS, Zarr, and HDF5, with checksummed files. Proprietary formats can be added by writing your own standardization algorithm.
- Can we use our own processing algorithms?
- Yes. Every standardization, check, and post-processing step can be replaced with your own algorithm, written in Python or as a container image and arranged alongside built-in algorithms in the same visual editor. AI coding tools and other assistants can also work the platform through MCP, the CLI, and SDKs.
- How can EmbodiFlow be deployed?
- Three plans are available. SaaS Team runs in the cloud with nothing to operate. Workstation is a GPU workstation with the platform preinstalled that runs fully offline. Enterprise runs on your private cloud for multiple teams, with high availability and custom features. Use Custom Configurations to size a deployment for your data.
- Which models can be trained on the platform?
- ACT, SmolVLA, OpenPI, and other policies can be trained on a frozen dataset version. Training can be watched as it runs, and the weights downloaded when it finishes. Every model traces back to its data, algorithms, and parameters.