Solarian Base
solarian is to Solarian Base what Next.js is to Vercel. You build a project
with solarian; Base stores its datasets, deploys its code, trains it on cloud
GPUs, tracks every run, serves its models and services, and lets you drive its
robots from a browser. Most
actions work from the solarian base CLI and from the web app alike.
New to Base? Follow the Quick Start: it takes one project from a deployment to a trained model, served on Base and called by the robot's skills, and a robot driven from the browser.
What you can do, and where
| Capability | CLI | Web app |
|---|---|---|
| Log in, API tokens | solarian base login, logout, whoami | Settings → API tokens |
| Projects | solarian base link --create | Projects |
| Upload and view datasets | Not available yet | Projects → New Dataset, Datasets |
| Computed datasets, ops, review | Not available yet (solarian dataset works on local copies) | A dataset's Versions |
| Deploy a project | solarian base deploy, manifest | Project → Deployments |
| Project secrets | solarian base env | Project → Environment variables |
| Train on cloud GPUs | solarian base train | Start training |
| Follow runs | solarian base runs | Runs |
| Robots, relays, controller sessions | Not available yet, except solarian run-robot on the robot | Robots |
| Checkpoints of training jobs | solarian base models ls | Models |
| Hosted services and policy servers | solarian base endpoints | Services & Policies, Models → Serve |
Every capability is also in the REST API. The interactive reference is at
/api/docs/ on your Base, and the schema at /api/schema/.
Concepts
- Name
Project- Description
The unit everything else belongs to: datasets, deployments, training jobs, runs and robots. A solarian project on disk links to one with
solarian base link.
- Name
Dataset- Description
An uploaded recording in LeRobot, UMI (Zarr) or Solarian format. Uploaded datasets never change.
- Name
Computed dataset- Description
A tree of versions over an uploaded Solarian dataset, each one an op applied to its parent: mirror, merge, split, track, delete episodes.
- Name
Deployment- Description
An immutable snapshot of the project's code, with the trainers and train configs it contains. Training jobs run against one.
- Name
Training job- Description
One run of
solarian modal-trainon a cloud GPU, launched by Base from a train config of the project or from a stock trainer.
- Name
Run- Description
The experiment record of a training: metrics history, console logs, tags, notes and a summary. A job opens one; so does any training that logs with the
solarian_basebackend.
- Name
Endpoint- Description
A service or policy server of the project running on a Modal GPU at a stable URL, from a conf with a
server:block and, for a policy, a checkpoint a training job wrote. The robot's connection file points the project's confs at it.
- Name
Robot- Description
A physical robot registered in a project. Base hosts a relay for it, so it can be driven from a controller session in the browser.
Where to go next
- Quick Start — deploy, train and run a robot, end to end.
- Getting Started — account, CLI login, linking a project, API tokens.
- Projects & Datasets — uploads, the viewer, computed datasets and review.
- Deployments & Secrets — shipping code and environment variables.
- Training & Runs — cloud training jobs and the runs they log.
- Services, Policies & Models — serving checkpoints and services on cloud GPUs for the robot's skills.
- Robots & Controller — connecting a robot, the managed relay and controller sessions.
- CLI Reference — every
solarian basecommand and flag. - Self-hosting — running Base on your own machine.