Phospho

Phospho

phospho · Coding · Marketing · Other

Phospho is an open source AI robotics toolkit built on phosphobot. You record robot datasets, train an action model and run it on real hardware from one local dashboard.

Interface preview of Phospho

About Phospho

What Is Phospho

Phospho is a small company split between Paris and San Francisco, backed by Y Combinator, that builds software for affordable robot arms. Its main product is phosphobot: an open source package that runs on your own computer and gives you a browser dashboard for driving a robot, recording demonstrations and training AI models from those recordings. The company also sells starter kits with hardware it has tested.

The product answers a practical problem. Research papers describe teaching an arm to stack blocks or pick up a cloth, but someone who buys an SO-101 arm usually has nothing between "the servos move" and "the arm does the task by itself". Phospho puts teleoperation, dataset capture, training and deployment in one place. According to phospho, more than 2,000 robots run phosphobot and users have trained over 1,000 models with it.

The limits are worth stating up front. This is a developer tool, not a consumer appliance. You install from a terminal, calibrate motors by hand and wire up cameras yourself. Hardware is sold separately, and the handiest extras, VR control and cloud training, sit behind the €35/month PRO plan.

Getting Started

  1. Install phosphobot with the one line command for macOS, Linux or Windows, then run phosphobot run and open localhost in a browser.
  2. Plug the arm into USB, power it on and walk through calibration so the software learns each joint's range.
  3. Attach cameras and record a handful of demonstrations with a keyboard, a gamepad, a leader arm or a Meta Quest headset.
  4. Choose a model such as ACT or π0 and start training. Datasets and checkpoints stay in the LeRobot format, so they can go straight to Hugging Face.
  5. Run the trained policy from the same dashboard and watch the arm repeat the motion.

Product Information

A quick look at Phospho's pricing, supported platforms, and performance.

Free PlanYes
Paid Plans€0 - €35/mo
PlatformmacOS, Linux, Windows
Developerphospho
CategoryCoding · Marketing · Other
Release DateJan 2025
Latest UpdatedSep 2026
Website Visits20.6K
Website Global Rank1.3M
API AvailabilityYes

Best for

The users, tasks, and scenarios where this tool fits best.

Users

  • Robotics hobbyists
  • Machine learning engineers
  • Students and research labs
  • Makers building a demo

Tasks

  • Recording imitation learning datasets
  • Training and fine tuning action models
  • Running the same policy on different arms
  • Adding vision to a cheap arm

Scenarios

  • A first robot project on a desk
  • Reproducing a paper or a tutorial
  • Teaching a class or workshop

Key features

Local Dashboard for Every Robot

phosphobot run starts a server on your machine and serves a dashboard at localhost. From there you move joints, check camera feeds, record episodes and launch training, all without a cloud login. Working locally also keeps raw robot footage on your own drive, which matters when a lab or client has rules about uploading data.

Dataset Recording With Four Input Methods

You can drive the robot with a keyboard, a gamepad, a leader arm or a Meta Quest headset. Each method writes the same episode format, including joint states, camera frames and task labels. Recording a usable batch of demonstrations usually takes a few short sessions, and the dashboard lets you review and delete bad takes before training.

One Click Action Model Training

The training panel covers ACT, π0 and gr00t-n1.5, three of the action models most used for imitation learning on small arms. You pick a dataset, a model and a number of steps, then watch loss in the browser. Checkpoints are saved in LeRobot format, so a model trained here still works with the wider LeRobot tooling.

VR Teleoperation on Meta Quest

With a Quest 2, Pro, 3 or 3s and the phospho teleoperation app, you control the arm with your hands and see through its cameras. VR makes bimanual work and fine placement easier than keyboard nudging, and it speeds up data collection. This feature needs the PRO subscription, which is the main reason hobbyists upgrade.

Works With the Arms People Actually Buy

Phospho supports the SO-100 and SO-101 arms, the Unitree Go2 quadruped and the Agilex Piper, and the open source core lets you add your own robot. Camera support covers standard USB webcams, depth sensors and stereo rigs. A single codebase across those devices saves the usual hunt for a driver or a fork that someone abandoned in 2024.

Open Source Core and Developer API

The repository ships under the MIT license, so you can read the control code, patch a driver and ship your own build. A local HTTP API lets scripts send commands and pull sensor readings, which is how most teams connect phosphobot to an existing pipeline. Python users get the same access without touching the dashboard.

Pros and cons

Pros

  • The free and open source tier is genuinely usable: control, recording and training all work without paying.
  • Setup is short for a robotics stack: one install command, a browser dashboard and no cloud account.
  • LeRobot and Hugging Face compatibility means datasets and models move into the wider robot learning world without conversions.
  • Support for several arms and camera types in one tool cuts the number of half finished scripts a team maintains.
  • The MIT license gives labs and startups room to modify the code and keep their changes private.

Cons

  • VR control and advanced training require the €35/month PRO plan, and the hosted app was paused when we checked, so those features depend on the local install and a stable connection to phospho's services.
  • The software is only half the cost: an arm, cameras and a computer add several hundred euros before the first dataset exists.
  • Calibration, USB drivers and camera framing still trip people up, and the docs assume you are comfortable in a terminal.
  • Documentation leans heavily on the SO-101, so owners of other robots will spend time reading source code and Discord threads.

Frequently asked questions

Yes. The core phosphobot software is open source under the MIT license, and controlling a robot, recording datasets and training models cost nothing. The paid PRO plan adds VR control, advanced training and support.