Research report · In Progress
Urban Segmentation for Delivery Robots
In-progress perception project · repository coming soon
This project frames semantic segmentation as a robotics safety module rather than a stand-alone vision demo: perception outputs should support navigability, obstacle awareness, and uncertainty-aware planning.
Abstract
Build a semantic-perception project for sidewalk and urban delivery robots, focused on navigability, obstacles, crossings, and scene reliability.
Scientific question: Which semantic scene cues are most useful for safe low-speed urban robot navigation?
Contribution: A planned segmentation pipeline connecting semantic perception to navigation-relevant risk features.
Marked research artifact slot
Perception-to-planning architecture
Architecture figure slot for verified diagrams of modules, data flow, and software boundaries.
Marked research artifact slot
Image · segmentation · traversability cost
Pipeline slot for GIFs or animations that explain estimation, planning, perception, or evaluation steps.
Marked research artifact slot
Synthetic semantic overlay
Demo slot reserved for real project videos or synthetic demos that are explicitly labeled as synthetic.
Research questions
- How can semantic labels be converted into navigation-relevant costs?
- How should segmentation uncertainty be exposed to the planner?
- Which failure cases matter for sidewalk delivery robots?
Methods
Implementation status
Implemented
- Research scope
- Project-page scaffold
Research prototype
- Synthetic segmentation overlay
Planned
- Dataset selection
- Baseline model
- Navigation-cost conversion
- Failure-case gallery
Experiments and metrics
Experiments
- Planned segmentation baseline
- Planned qualitative overlay inspection
Metrics
- mIoU — planned
- class-wise error — planned
- navigation-cost disagreement — planned
Quantitative benchmark tables will be added only after reproducible experiments are available.
Limitations
- No public repository is linked yet.
- Quantitative claims are intentionally withheld until reproducible benchmark runs are available.
- Visual figures and GIFs are placeholders unless a project page explicitly states that an artifact is generated from real experiments.
Reproducibility plan
Experiment workflow
Each module is prepared to document configuration files, datasets, evaluation metrics, and repeated experiment runs before quantitative claims are shown.
Repository setup
Public repositories are linked when available. Private, planned, or incomplete repositories remain clearly labelled rather than being presented as finished systems.
Literature context
- Semantic segmentation
- Traversability estimation
- Urban delivery robots
Roadmap
- Create repository
- Add reproducible dataset notes
- Connect to DynNav risk maps