From raw recordings to deployed policies
EmbodiFlow turns robot teleoperation, human egocentric and UMI recordings into trusted dataset versions through programmable workflows, then takes them through training, evaluation and deployment.
On AWS cloud, on your GPU workstation, or on your private cloud: the same product.
Sample workflow run · grasp-cups / rec-0142
Ingest MCAP
Quality gate
Standardize to LeRobot v3
Face anonymization
Stereo depth
AI episode suggestions
Recording hours are not usable hours
The bottleneck in embodied data is not how much you record, but how much you can use, trust and reproduce.
Scattered scripts
Conversion, QA and post-processing scripts live on laptops with no versions. A new person or a new batch breaks them.
Formats that do not agree
ROS bags, MCAP, vendor HDF5, GoPro video and headset captures differ in units, frames and action meaning.
Quality nobody can state
Dropped frames, broken timestamps and unsynchronized streams surface only when training fails to converge.
Results you cannot reproduce
Which data, which script version and which parameters trained this model? Most teams cannot answer.
One platform, the whole loop
Every step between a recording and a deployed policy is bound to an immutable version, and every processing step can be your own algorithm.
Your AI assistant drives every step through MCP and the CLI
Built for embodied data, end to end
Your workflow, your algorithms
Every step from raw recording to training data is a replaceable algorithm. Use the built-in ones, or swap in your own standardizer, quality check or perception model.
Write a few lines of Python, try it on a few recordings, save it as a version — or push a container image.
Register a private recorder format as a team format; uploads are standardized automatically.
Built-in and custom algorithms share one contract and one visual editor.
The platform schedules GPUs, parallelism, retries and caching. You write the logic.
Built-in AI, and open to yours
Built-in AI does the repetitive data work. Through MCP, the CLI and SDKs, assistants such as Claude and Cursor can do anything you can do on the platform.
AI suggests episode boundaries, subtasks and skills. A person confirms before anything takes effect.
Find episodes in natural language, or by what they look like and how they move.
A data map shows clusters and gaps across every episode in a project.
Assistants use your access token: never more permission than you, with the same audit trail.
Data, training, evaluation and deployment in one place
Start training from a dataset version, compare models offline, and deploy as a cloud inference service or an edge package on the robot.
Mainstream policy and VLA model families, from ACT to SmolVLA and OpenPI.
Offline evaluation compares predicted and recorded actions against a baseline.
Works with both the OpenPI and LeRobot official clients — no changes on the robot.
Teleoperation, human egocentric and UMI, treated alike
Every input is standardized to LeRobot v3 on a 30 Hz time grid, so different recorders can be trained together.
Human egocentric data is first-class: camera and IMU only, with derived actions.
A complete UMI chain: GoPro telemetry, fisheye calibration, SLAM mapping and gripper state.
Stereo depth, hand pose, camera pose and face anonymization built in.
If it was not checked, it never shows as passed
Quality gates sit at the entrance and keep untrainable data out, at the level of individual files and episodes.
Six statuses: pass, warning, fail, error, not evaluated and not applicable. The last three are never green.
Drop only the bad episodes, not the whole dataset.
Every finding carries its threshold and evidence interval, and jumps to that moment in replay.
Every model can say what made it
Dataset versions are frozen. Lineage runs from raw recording through processing, quality, version, export and training to the model.
Each run freezes the algorithm version, image, parameters and inputs.
Exports record a checksum for every file; regenerating must produce identical bytes.
New data never changes a historical version.
Sign up and go, or run it on your own GPUs
The same product runs on AWS cloud, fully offline on a single GPU workstation, or on your private cloud cluster. Self-hosted data never leaves the building.
Workstation ships as a Dell GPU workstation with everything preinstalled; Enterprise runs multi-node with high availability.
Offline license verification, built-in object storage, continuous backup and point-in-time restore.
When a license expires the platform becomes read-only. Your data stays readable and exportable.
No lock-in on hardware or formats
Export standard LeRobot v3, RLDS, Zarr or HDF5. Point huggingface_hub and LeRobot at the platform with one environment variable.
Full REST API with OpenAPI, TypeScript and Python SDKs, CLI and MCP server.
Webhooks notify you when work finishes — no polling.
Everything you can click, your assistant can call
MCP tools are generated from the public API, so new features reach your assistant immediately. Write operations take idempotency keys and failures return named reasons.
"Upload this folder to the grasping project and tell me the quality once it is processed."
"Write a quality check that the gripper stays visible the whole time."
"Make a version from the passing episodes, train ACT, and compare it with the last model."
$ embodiflow auth login
$ embodiflow projects list --output json
$ embodiflow uploads create --project-id <project> --source-path ./recordings
$ embodiflow versions download <team>/<project>@<n>Compared with how it is done today
Home-grown scripts
No versions, hard to reuse
Algorithms registered as versions, visual workflows, scheduling and lineage
Closed all-in-one platforms
Fixed steps, no room for your algorithms
Every step can be your own algorithm
Web-only platforms
AI assistants cannot use them
MCP, CLI and SDKs open everything to your assistant
File names and manual review
Breaks down at scale
Semantic search, similarity and a data map
Separate tools for data, training and deployment
Manual hand-offs between each stage
One platform, unbroken lineage
Robot data visualization tools
You can look, but not build datasets
Replay to find problems, then process, version and export
General labeling and curation tools
No grasp of robot time series or actions
Robot-specific quality rules and action semantics built in
| Current practice | Limitation | EmbodiFlow |
|---|---|---|
| Home-grown scripts | No versions, hard to reuse | Algorithms registered as versions, visual workflows, scheduling and lineage |
| Closed all-in-one platforms | Fixed steps, no room for your algorithms | Every step can be your own algorithm |
| Web-only platforms | AI assistants cannot use them | MCP, CLI and SDKs open everything to your assistant |
| File names and manual review | Breaks down at scale | Semantic search, similarity and a data map |
| Separate tools for data, training and deployment | Manual hand-offs between each stage | One platform, unbroken lineage |
| Robot data visualization tools | You can look, but not build datasets | Replay to find problems, then process, version and export |
| General labeling and curation tools | No grasp of robot time series or actions | Robot-specific quality rules and action semantics built in |
Three ways to run it
SaaS Team
One team, hosted on AWS cloud. Nothing to operate, always up to date.
Workstation
A Dell workstation with GPUs and EmbodiFlow preinstalled. Unbox it in your building and start, fully offline.
Enterprise
Multiple teams on your private cloud cluster, with high availability and custom features.