Real robots. Real data. Real constraints.

The physical world is very good at finding assumptions.

At ko-br, a small robotics startup, I get to move between very different jobs. Some days I'm figuring out why a camera stream is dropping frames; other days I'm wiring a lead arm, looking into a robot-learning paper, or testing a digital twin in Isaac Sim. I like that range.

With UR10e arms and multiple Intel RealSense D435i cameras, small details matter: USB 2 versus 3.2, dropped frames, depth noise, calibration and cables that look fine until you start recording. Getting a stream to run is one thing; getting consistent data from it is the real work.

From soldering to simulation.

I've soldered and wired encoders for a lead arm designed to capture demonstrations, then switched contexts to NVIDIA Isaac Sim and digital-twin experiments with our UR10e. Those are very different tools, but they're connected by the same question: how do we get useful behaviour from real-world robot data?

I'm also studying robot-learning approaches such as Physical Intelligence's π₀. That work is research and experimentation for now, not a claim that I've trained a production robot policy.

Another view from Universal Robots.

I joined Universal Robots in Munich in August 2026. It's a different perspective from startup R&D, and I'm learning a lot from working closer to industrial robotics. Together, these experiences are pulling me further toward embodied AI—the meeting point of sensing, learning and action.

Technical threads

• Robot-mounted RGB-D sensing, camera configurations and depth-quality evaluation.
• Encoder integration, lead-arm hardware and demonstration data.
• ROS/Webots navigation experiments with LiDAR and depth sensing.
• Exploration of simulation-to-real workflows and robot-learning architectures.