Çoban — A Robot Arm Driven by Turkish Commands
A closed-loop robot arm that takes commands in Turkish and verifies every step with a camera: I adapted the Shepherd Robotics (YC F26) approach to Turkish industry. 5 scenarios with the real Gemini API, 43 automated tests.
- Period
- 2026
- Stack
- Python · Gemini API · Digital twin (simulation)
The question
Can a small manufacturer in Türkiye tell a robot to "scrap the cracked connector" without a robot programmer? The starting point was the "new operating systems for the physical world" call in Y Combinator's Fall 2026 Request for Startups list, along with Shepherd Robotics, which was accepted into YC in the same batch. Çoban is a clone of that approach adapted to Turkish industry.
Components
- Closed-loop task execution: At every step, Gemini looks at the fresh camera image to choose the next move and checks whether the previous step actually succeeded.
- Vision + touch fusion: The gripper sensor catches 1–2 cm slips that can't be seen in the image, and the system replans.
- Completion check: When the model says "done," a separate quality control step counts what's in the final image; if anything is missing, the task continues.
- Digital twin: A simulation that runs the same code without a robot, with wire harness kitting and cracked-part quality control scenarios modeled on automotive suppliers in Bursa.
- KVKK-compliant image processing: Workers' faces are blurred on the device before any image is sent to the cloud.
- Data flywheel: Every task is recorded as a dataset for training an open-source robot model in the future.
Results with the real Gemini API
| Task | Result | What it shows |
|---|---|---|
| Put the red cubes in the yellow box | Completed (5 steps) | Multi-step planning, with every step verified against the image |
| Throw the cracked connector in the scrap bin | Completed (3 steps) | Visual quality control |
| Put the purple cube on the white tray | Impossible | Not moving the robot at all for an object that isn't on the table |
| Same task, with a 40% grip failure rate | Completed (8 steps) | The sensor catches 2 slips; a premature "done" is rejected |
The system was tested in 5 scenarios with the real Gemini API and validated with 43 automated tests.
What taught me the most: the failures
The model missed a slipping part, mistook a white connector for a black one, and said "done" before the task was finished. I closed each of these with an engineering fix. The model saying "done" doesn't mean the job is done.
The goal
A low-cost, reliable automation approach that understands Turkish, for a Turkish manufacturing sector whose robot density is below the world average.
I would like to thank Asst. Prof. Dr. Nurettin Şenyer for the guidance and academic contributions.