I help robotics startups turn unstable prototypes into systems that can be deployed and sold to customers

→ robot oscillated near the dock
→ vision pipeline produced desynchronized data

→ two-stage docking: coarse and precise
→ separate acquisition from perception

→ industry standard (99.9% success rate)
→ stable FPS (no drops over time)

→ factory could not be mapped reliably for production deployment
→ inconsistent maps prevented stable robot localization across the facility

→ customized SLAM for sparse-loop industrial environments
→ improved trajectory alignment and constraints handling for large-scale mapping

→ stable factory-scale maps suitable for real deployments
→ enabled long-term autonomous operation inside an industrial facility

⚙️ Custom pose estimation for robotic manipulator when the standard stack was not enough

Problem: The robot pose could not be estimated reliably with an off-the-shelf approach.

What I did: I implemented a custom optimization-based method that combined camera observations, robot kinematics constraints, and motion assumptions into one pose estimation problem.

When this matters: your project is research-heavy, the standard robotics stack does not fit your environment, and you need someone who can implement algorithms... not only integrate existing packages.

🗺️ Multi-session SLAM and map stitching

Problem: A single mapping run was not enough.The robot needed to combine multiple mapping sessions from different rosbags, trajectories, and operating conditions.

What I did: I customized the SLAM workflow to load previous mapping state, freeze existing trajectories, add new ones, and align them using inter-trajectory constraints.

When this matters: mapping has to become a repeatable deployment process. Not a demo that only worked once.

🧭 Sensor fusion failures caused by clock drift

Problem: The autonomy stack looked correct, calibration was verified, and sensor data appeared valid. The real issue was that different sensors were operating on slightly different clocks, causing subtle timing errors throughout the autonomy stack.

What I did: I implemented a synchronization architecture based on GNSS, PTP, Chrony, and ROS2 to establish a single time reference across all sensors and compute devices. The fix was not a better localization algorithm. The fix was making every component agree on exactly when an event happened.

When this matters: you keep tuning localization or sensor fusion, but the real problem is hidden in timestamps.