LABS (Learning & Acquisition of Behaviors Suite) is an open-source, microservice-driven platform for robotics data collection and Physical AI research. It coordinates actuators, sensors, and teleoperation devices in a unified pipeline — from demonstration to dataset export.
From teleoperation with GELLO to episode recording to LeRobot-format dataset export — LABS covers the entire data pipeline on one platform, reducing toolchain fragmentation.
MODULAR, CONTAINERIZED ARCHITECTURE
Every component — robot controller, camera driver, data recorder — runs as an isolated microservice via Docker Compose, making it easy to swap hardware or extend with custom services.
NATIVE FR3 DUO INTEGRATION
Designed alongside FR3 Duo. Supports dual-arm configurations out of the box, with native drivers for Franka FR3, GELLO, Robotiq grippers, RealSense, and ZED cameras.
STANDARDIZED, REUSABLE DATASETS
Episodes are stored in MCAP format and exported as LeRobot v2.1 datasets. Reproducible hardware configs let labs share and reuse data across institutions.
WHAT IS INCLUDED
THREE PIPELINE SERVICES, ONE STACK
Data Collection FastAPI state machine orchestrator that manages teleoperation mode, episode recording, and workflow transitions via a web UI.
Data Recorder Captures all active ROS 2 topics and saves episodes in MCAP format, synchronized across all connected actuators and sensors.
Data Processor Post-processes and transforms raw episodes, applies FFmpeg video encoding, and exports to LeRobot v2.1 format for model training.
FROM SETUP TO DATASET IN FIVE STEPS
LABS is designed to minimize time between hardware setup and the first recorded episode.
Define your hardware layout in a deployment config file — robot IPs, sensor types, gripper models.
3. COLLECT EPISODES VIA THE UI
Open the Data Collection UI at localhost:4000, select a task, start teleoperation, and record demonstrations.
2. START THE SOFTWARE STACK
Run task start from your station folder. Tilt builds all container images and launches the full service mesh.
4. EPISODES SAVED AUTOMATICALLY IN MCAP
The data recorder captures and persists episodes in real time. Review them directly in the Episodes tab of the UI.
5. EXPORT TO LeRobot DATASET
Run the data processor to convert your MCAP episodes into a standardized LeRobot v2.1 dataset ready for model training.
PROTOTYPE PRODUCTS
THE REFERENCE PLATFORM FOR PHYSICAL AI
A new generation of robot systems designed to enable physical intelligence research, from tactile teleoperation to mobile manipulation. These systems enable physical AI activities, including data collection at scale and real-world learning.
The resource hub for developers working with the Franka Research 3 robot platform. Dive into a wealth of tools, libraries and documentation designed to accelerate your development journey. Whether you're a seasoned developer or just starting out, you'll find everything you need to build, create and innovate.
-1.jpg)
_resized.png)
