Project Overview
This thesis project at Sevensense Robotics focused on pallet localization for docking applications of autonomous mobile robots in indoor industrial environments. The system was designed around grayscale vision, modularity, and robustness rather than a narrow lab-only setup.
My work included building a 6-DoF pose estimation module together with a multi-object pose tracking module. A major part of the effort was not just creating the proof of concept, but testing it extensively, identifying shortcomings, and improving the system to make it more reliable in practice.
Technical Architecture
Grayscale Perception
Built around grayscale imagery for an industrial robotics use case where reliability and simplicity mattered
6-DoF Pose Estimation
Estimated pallet pose accurately enough to support autonomous docking behavior
Multi-Object Tracking
Tracked relevant targets across frames to stabilize the overall localization pipeline
Robustness Testing
Used extensive testing to expose weak points and iteratively improve the proof of concept
Why the Problem Was Interesting
Pallet localization sounds narrow on paper, but it brings together several core robotics challenges:
- Pose quality matters: Small geometric errors can directly affect docking behavior and operational reliability
- Industrial scenes are messy: The perception pipeline had to cope with realistic variation rather than curated examples
- Modularity helps iteration: A modular design made it easier to improve individual parts of the stack
- Tracking complements detection: Stable downstream behavior depends on both recognition and temporal consistency
Implementation Focus
Pose estimation module: A central part of the project was estimating pallet pose with enough accuracy and repeatability for docking applications, which meant taking the geometric side of the problem seriously.
Tracking layer: I also developed a multi-object pose tracking component to keep the system stable over time and reduce sensitivity to noisy single-frame outputs.
Robustness-driven iteration: Extensive testing surfaced practical weaknesses in the proof of concept, which then guided improvements to make the overall pipeline more dependable.
Impact
This project strengthened my experience in 3D computer vision, pose reasoning, and how to evaluate a system beyond the happy path.
It also reflected a pattern that shows up throughout my work: useful robotics perception is rarely about one algorithm alone. It depends on how estimation, tracking, testing, and system design fit together.
Related Open-Source Work
Some of the pallet-focused work I share publicly is linked below. The production and thesis systems themselves were developed in a research and company context, so the public code should be viewed as related work rather than a full project dump.