phospho-app/phosphobot

Control AI robots. Community-driven UI middleware for controlling robots, recording datasets, training action models. Compatible with SO-100 and SO-101

What it solves

Phospho provides a unified software stack to bridge the gap between physical robot hardware and AI models. It simplifies the complex process of recording movement data (teleoperation), training Vision-Language-Action (VLA) models, and deploying those models back onto robots for autonomous control.

How it works

The system operates as a server with a web-based dashboard that connects to various robot hardwares and cameras. Users can control robots via a keyboard, gamepad, leader arm, or Meta Quest headset to record datasets of specific tasks. This data is then used to train action models (such as ACT, smolVLA, π0.5, or gr00t-n1.5) either through the webapp or custom code. Once trained, models are exported to Hugging Face and can be used to drive the robot autonomously via the dashboard or a Python API.

Who it’s for

It is designed for robotics developers and AI researchers who want to quickly collect data and train VLA models without building the entire infrastructure from scratch, as well as users of supported robots like the SO-100, SO-101, and Unitree Go2.

Highlights

  • Multi-device Teleoperation: Supports control via headsets, gamepads, keyboards, and leader arms.
  • VLA Model Integration: One-click training for models like ACT and gr00t-n1.5.
  • Broad Hardware Support: Compatible with various robots (SO-100, Unitree Go2, AgileX Piper) and most camera types (depth, stereo, classic).
  • Ecosystem Compatibility: Fully compatible with LeRobot and Hugging Face.
  • Simulation Environment: Includes a fully accessible simulation for testing without physical hardware.

Related

  • Project
    visose/Robots

    A plugin for Rhino 8 and Grasshopper that allows users to create, simulate, and generate manufacturer-specific code for various industrial robots.

  • Project
    amazon-far/abc

    A scalable behavior cloning framework for robotics that provides Diffusion Transformer (DiT) and Vision-Language-Action (VLA) models for training and deploying robot policies.

  • Project
    songge8/CF-Drone

    An open-source flight-control firmware for ESP32-family microcontrollers (ESP32, ESP32-S3, ESP32-C3) that flies small drones like the Chenguang E1 series. It handles attitude sensing, motor PWM control, low-battery safety, LED status, and a browser-based remote control — all configurable through a debug console.

  • Project
    BeamNG/BeamNGpy

    A Python API for BeamNG.tech that enables remote control of vehicle simulations, sensor data collection, and AI-driven vehicle behavior for research and robotics.