XR Robot Teleop Server

A high-performance Python server designed to stream 360° panoramic video and full-body tracking data to XR headsets (like Quest 3, Vision Pro) for immersive robot teleoperation. It uses WebRTC for low-latency communication and features a modular architecture for easy customization.

Key Features

• Low-Latency Streaming: Built on aiortc and FastAPI for efficient WebRTC communication.
• 360° Video Reprojection: Dynamically transforms 360° equirectangular video into a perspective view based on the user's head orientation, providing an immersive FPV experience.
• Full Body Tracking: Receives, deserializes, and processes full-body, upper-body, and hand skeleton data from XR clients (e.g., a Unity app).
• Modular & Extensible: Easily replace video sources (e.g., from a file, a live camera, or a GoPro stream) and video transformations with your own custom classes.
• Hardware Acceleration: Utilizes FFmpeg's hardware acceleration capabilities (videotoolbox, d3d11va, vaapi, etc.) for efficient video decoding, minimizing CPU load.
• Coordinate System Handling: Includes utilities to convert coordinate systems between different platforms (e.g., Unity's left-handed Y-up to a standard right-handed Z-up).
• Rich Data Schemas: Provides clear Python enumerations for OpenXR and OVR skeleton bone IDs, simplifying data interpretation.
• Integrated Visualization: Comes with an example to visualize the received body pose data in 3D using the Rerun SDK.

System Architecture

The server acts as a bridge between an XR client (e.g., a VR headset running a Unity application) and a robot control system. The client sends head pose and body tracking data to the server, which in turn processes it and sends back a reprojected video stream. The processed body pose can then be used to command a robot or a simulation.

graph TD
    subgraph "XR Client (e.g., Unity on Headset)"
        A[User Input: Head Pose + Body Tracking] --> B{WebRTC Connection};
        B --> C["Send Data Channels (body_pose, camera)"];
        D[Receive Video Stream] --> E[Display to User];
        B --> D;
    end

    subgraph "Python Server (xr-robot-teleop-server)"
        F{WebRTCServer};
        G[Data Channel Handlers];
        H[Video Pipeline];

        F <--> G;
        F <--> H;

        subgraph G [Data Channel Handlers]
            G1["'body_pose' handler"];
            G2["'camera' handler"];
            G1 --> G3[Deserialize Pose Data];
            G2 --> G4[Update Head Pose State];
        end

        subgraph H [Video Pipeline]
            H1[VideoSource: 360° Video];
            H2[VideoTransform: Reprojection];
            H3[Reprojected Video Track];
            H1 --> H2;
            H2 --> H3;
        end

        G4 --> H2;
    end

    subgraph "Robot Control System (Application Layer)"
        I["Robot/Simulation (e.g. MuJoCo)"];
        G3 --> I;
    end


    C --> F;
    H3 --> D;

    style XR Client fill:#d4fcd7,stroke:#333,stroke-width:2px
    style Python Server fill:#cde4f7,stroke:#333,stroke-width:2px
    style Robot Control System fill:#f7e8cd,stroke:#333,stroke-width:2px

Getting Started

Installation

pip install xr-robot-teleop-server

Development

1. Clone the repository:

   git clone https://github.com/your-username/xr-robot-teleop-server.git
   cd xr-robot-teleop-server
   

2. Install the package: For basic usage, install the package in editable mode:

   pip install -e .
   

   To include dependencies for the visualization example (rerun), install with the [viz] extra:

   pip install -e ".[viz]"
   

Running the Examples

This project includes examples to demonstrate its capabilities.

Record and Visualize Full Body Pose

This example starts a WebRTC server that listens for full-body tracking data from a client, deserializes it, and visualizes the skeleton in 3D using Rerun.

1. Run the server:

   python examples/record_full_body_pose.py --visualize
   

   The server will start and print: INFO Starting xr-robot-teleop-server...

2. Connect your client: Connect your XR client to the server at http://<your-server-ip>:8080. Once the WebRTC connection is established, the server will start receiving data.

3. Visualize: A Rerun viewer window will spawn, displaying the received skeleton poses in real-time.

   

Usage

The core of the server is the WebRTCServer class. You can instantiate it with factories for creating video tracks and handlers for data channels.

from xr_robot_teleop_server.streaming import WebRTCServer
from xr_robot_teleop_server.sources import FFmpegFileSource
from xr_robot_teleop_server.transforms import EquilibEqui2Pers
from aiortc import MediaStreamTrack
import numpy as np

# 1. Define a shared state object (optional)
# This object is shared between the video track and data channel handlers
# for a single peer connection.
class AppState:
    def __init__(self):
        self.pitch = 0.0
        self.yaw = 0.0
        self.roll = 0.0

# 2. Define a handler for an incoming data channel
def on_camera_message(message: str, state: AppState):
    # Parse message and update the shared state
    data = json.loads(message)
    state.pitch = np.deg2rad(float(data.get("pitch", 0.0)))
    state.yaw = np.deg2rad(float(data.get("yaw", 0.0)))

# 3. Define a factory for the video track
# This factory creates the video source, the transform, and the track itself.
class ReprojectionTrack(MediaStreamTrack):
    kind = "video"

    def __init__(self, state: AppState):
        super().__init__()
        self.state = state
        self.source = FFmpegFileSource("path/to/your/360_video.mp4")
        self.transform = EquilibEqui2Pers(output_width=1280, output_height=720, fov_x=90.0)

    async def recv(self):
        equi_frame_rgb = next(self.source)
        rot = {"pitch": self.state.pitch, "yaw": self.state.yaw, "roll": self.state.roll}
        perspective_frame = self.transform.transform(frame=equi_frame_rgb, rot=rot)
        # ... create and return av.VideoFrame ...

# 4. Configure and run the server
if __name__ == "__main__":
    data_handlers = {
        "camera": on_camera_message,
        # Add other handlers for "body_pose", "left_hand", etc.
    }

    server = WebRTCServer(
        state_factory=AppState,
        video_track_factory=ReprojectionTrack,
        datachannel_handlers=data_handlers,
    )

    server.run()

Project Structure

.
├── examples/                 # Example scripts showing how to use the server
├── src/
│   └── xr_robot_teleop_server/
│       ├── schemas/          # Data schemas for skeletons (OpenXR) and poses
│       ├── sources/          # Pluggable video sources (FFmpeg, OpenCV)
│       ├── streaming/        # Core WebRTC server logic
│       └── transforms/       # Pluggable video transformations (e.g., reprojection)
├── tests/                    # Unit and integration tests
└── pyproject.toml            # Project metadata and dependencies

Dependencies

• Core: aiortc, fastapi, uvicorn, numpy, loguru
• Transforms: equilib
• Video Sources: opencv-python (optional), ffmpeg (must be in system PATH for FFmpegFileSource)
• Visualization: rerun-sdk

See pyproject.toml for detailed dependency information.

License

This project is licensed under the terms of the LICENSE file.