Project IRA is a team project at the University of Tübingen with the goal of designing and deploying an interactive robotic arm pipeline.
IRA is a full robotic pipeline consisting of a GUI, webserver setup, manual robot controls, movement/dataset recording scripts, motion playback, model training scripts, asynchronous model inference and more.
Built on the leRobot SO101, IRA is a highly accessible starter project for robot enthusiasts, students, schools and universities alike to step foot into the world of VLA robotics.
IRA is built as a practical desk assistant for everyday use or even for game nights.
Built with 4 different modes:
- Desk Cleanup
- Lego Sorter
- Dice Roller
- Play Fetch
All tasks are included in our dataset available at Huggingface.
The complete dataset is leRobot compliant and can be trained on as is.
IRA is available in two versions:
- SmolVLA
- Pi0.5
Choose your model based on your system specs (see the Pipeline Documentation for details). The IRA dataset can also be used to train any other open-weight model you'd like.
Our pretrained models and their respective checkpoints are also available at our Huggingface organization page. Please follow the instructions on huggingface regarding all models and datasets.
To run the complete system (GUI + robot + VLA together), follow our Quick Start Guide.
For the full pipeline (building the robot, recording datasets, finetuning and running inference), refer to our Pipeline Documentation. For pictures of the physical build and the hardware parts, see the Build Setup Documentation.
Our documentation is available here:
Full walkthrough:
Additional documentation:
Github Repo specific additional documentation:
Refer to this documentation for any further instructions regarding the functionality and setup of Project IRA.