LeRobot gets you started. Neuracore gets you to production.
LeRobot put imitation learning within reach of anyone with a low-cost arm, which is a useful thing to have done. It is a good place to learn and prototype. Once you move past a bench and a 3D-printed arm, a managed platform is the better tool. This page maps where each one fits.
Where LeRobot fits.
LeRobot is an open-source library from Hugging Face that lowered the barrier to entry for real-world imitation learning: pretrained policies, shared datasets, simulation environments, and support for low-cost, accessible hardware. For learning the field or prototyping a single task on one or two cheap arms, it is a good place to start.
Open and accessible
Free, open source, and community-driven, with a fast-moving ecosystem of models and shared datasets on the Hub.
Great for learning
One of the best on-ramps into imitation learning and real-world robotics for students, researchers, and tinkerers.
Low-cost hardware
First-class support for inexpensive, accessible arms makes it easy to go hands-on without a big hardware budget.
- ↳You are learning or teaching imitation learning.
- ↳You are prototyping a single task on one or two low-cost arms.
- ↳You want full control of the code and are happy to self-host.
- ↳Your project is research-first, where reproducibility matters more than uptime.
- You are moving from a demo to a real deployment or a fleet.
- Algorithm performance and reliability directly affect the business.
- You need managed scale: multi-GPU training and multi-robot data, without running the infra.
- You want one loop from collection to deployment, with support behind it.
From first demo to a production fleet.
| Dimension | LeRobot | Neuracore |
|---|---|---|
| Best fit | Research, education, prototyping | Production teams and organizations at scale |
| Infrastructure | Self-hosted; you run and maintain it | Managed cloud or private / self-hosted, your choice |
| Data pipeline | Datasets and tooling you assemble yourself | Streaming, sync, curation, and versioning built in |
| Training | Bring your own GPUs and orchestration | One-call multi-GPU runs, autotuned, in the cloud or on your own hardware |
| Deployment | DIY serving and control integration | One-line deploy with real-time closed-loop inference, cloud or local |
| Scale & support | Community forums and issues | Multi-robot scale with a team and SLAs behind you |
LeRobot is an open-source project and evolves quickly; this reflects the typical self-hosted experience. You can also bring LeRobot-format datasets into Neuracore, so the two are not mutually exclusive.
Neuracore closes the flywheel.
A library gives you the pieces. A platform connects them into one loop, where deployment failures flow back as data and every iteration is faster than the last. That loop is where production robotics is won.
Collect
Demonstrations stream in from every rig, live and durable.
Curate
Clean, synchronized, versioned datasets, ready to train.
Train
State-of-the-art policies on managed GPUs.
Deploy
Failures trace back to data and feed the next run.
Outgrowing your own stack?
If you started on LeRobot and are hitting scale, reliability, or deployment walls, that is exactly the handoff we are built for. Bring your datasets.