GitHub  •  LinkedIn  •  X  •  Discord

Telekinesis Data Engine

Telekinesis Data Engine turns real-world, online-collected robot and perception data into standard, trainable datasets within the Telekinesis ecosystem.

Features

It includes:

• MCAP logging: capture every publisher on the Zenoh network into a single self-describing .mcap file, and read it back into live objects

• Detection-dataset logging: write labelled frames straight to a trainable YOLO (Ultralytics) or RF-DETR (COCO) dataset, with automatic train/val/test routing

• LeRobot logging: record episodic robot data—including states, actions, tasks, images, and videos—into LeRobot datasets, with support for creating, overwriting, and resuming datasets

• Decoupled pub/sub writers so disk I/O never stalls the capture loop

• FiftyOne-based visualization of detection datasets

Requirements

Current support for Python versions - 3.10, 3.11, 3.12.

Note

• Telekinesis Data Engine is currently in its early development phase (pre-1.0). hello
• There will be continuous version updates which introduce new features and optimizations. To have the latest features, please always install or upgrade to the latest version of the package.

Installation

pip install telekinesis-dataengine

Example

Run a sample python code to quickly test your installation.

1. Create a Python file named dataengine_example.py in a directory of your choice, and copy paste the below:

   import numpy as np
   from telekinesis.dataengine import DetectionLogger
   
   # Create a logger that writes an Ultralytics YOLO dataset.
   # categories=None builds the class table dynamically from what you log.
   logger = DetectionLogger.create("yolo", "results/demo_dataset", mode="overwrite")
   
   # A labelled frame: a raw image (HxWxC) + COCO-style boxes [x, y, w, h].
   image = np.zeros((480, 640, 3), dtype=np.uint8)
   annotations = [{"category_id": 0, "bbox": [100, 120, 80, 60]}]
   
   logger.log(image, annotations)   # auto-routed 80/10/10 across train/val/test
   logger.close()                   # writes images/, labels/, and data.yaml
   
   print("Dataset written to results/demo_dataset")
   

2. On a terminal, navigate to the directory where dataengine_example.py was created, and run:

   python dataengine_example.py
   

   Expected output:

   Dataset written to results/demo_dataset
   

   Swap "yolo" → "rfdetr" to emit a COCO dataset instead — the call sites stay the same.

You are now set up with Telekinesis Data Engine.

Resources

• Examples Runnable usage examples for the Telekinesis SDK: Telekinesis Examples

• Documentation Full SDK documentation and usage details: Telekinesis Documentation

• Development Building, testing, and contributing to this package: DEVELOPMENT.md

Support

For issues and questions:

• Create an issue in the GitLab repository.
• Contact the Telekinesis development team at support@telekinesis.ai or on Discord.