Mehmet Baha Dursun · Project Lead, Saha Robotik
2026 - Present · Early-stage project at Saha Robotik
In parallel, I am preparing the collaboration between Saha Robotik and SDU Robotics and helping shape the Eurostars proposal around this project. My goal is to connect the company project with my MSc thesis so that the research, engineering work, and European collaboration can grow together.
My academic supervisor, Associate Professor Iñigo Iturrate, described this direction in the recommendation letter supporting my thesis funding application: the letter records the SDU and Saha dialogue, the expected European grant proposal, and my plan to build AI for mobile manipulators in the restaurant industry.
Read the Recommendation Letter• Eurostars• Supervisor Profile
This is a new, early-stage project. The goal is a mobile, bimanual robot that can move through a restaurant, stop at a table, and collect plates, glasses, cutlery, and trays. Saha's existing mobile-robot stack will handle navigation and positioning, while the manipulation system will learn the table work from human demonstrations.
OpenArm was selected as the initial reference and prototyping platform so we can move quickly through hardware bring-up, teleoperation, data collection, and policy experiments. It is not the final product goal. As the project matures, the plan is to design Saha's own bimanual arm system and transition the learning pipeline to that hardware.
I take the work of Sunday Robotics and the Physical Intelligence lab as my main benchmark for this direction. As the final target, I want to follow this class of pipeline end to end and build a similar system for Saha's restaurant and service-robot setting, using our own mobile base, OpenArm hardware, and deployment experience.
I have taken responsibility for the complete project pipeline at Saha Robotik, including the research, planning, initial OpenArm setup, teleoperation, data collection, model pipeline, evaluation plan, and future integration with the company's mobile robots. I am leading both the technical work and the overall direction of the project.
I am starting from the lessons of ALOHA and UMI and working toward a scalable data and learning pipeline for the company. The longer-term plan includes glove-based human demonstrations, following the UMI direction and the kind of data pipeline shown by Sunday Robotics. I plan to build on a foundation model such as pi0.5 and to keep improving from there. There is a lot to learn, and there is no production system or autonomous table-clearing result yet; I have only recently started building the system and receiving the OpenArm hardware, so it is still early, but I believe strongly in the direction.
ALOHA provides the practical starting point for low-cost bimanual teleoperation and learning from demonstrations. Mobile ALOHA shows how that idea extends to a mobile platform and whole-body tasks. I use these projects as system-design references, while keeping Saha's existing navigation stack responsible for the first mobile deployments.
UMI is the reference for collecting portable, in-the-wild human demonstrations before retargeting them to a robot. The first Saha dataset will still come from the real OpenArm teleoperation path. Once that contract is proven, an UMI-style wearable path can expand data collection across real restaurant layouts.
Sunday's Memo demonstrates the product experience I care about: clearing tables, handling plates and delicate glasses, and learning from human demonstrations collected away from the robot. For Saha, the target is a restaurant-focused version of this task class, integrated with a commercially deployed service-robot base.
OpenArm is the initial physical reference platform, chosen to accelerate the first experiments rather than define the final product. LeRobot supplies the dataset, robot, and policy interfaces, while the current VLA path uses pi0.5 for language-conditioned action chunks. The same learning pipeline is intended to move to Saha's own future bimanual hardware.