Inspiration
My Roborock sometimes gets stuck under furniture, spends time on areas that already look clean, and uses up its battery before reaching dirtier spots. I wanted an extra pair of eyes to check what it had actually left behind.
That frustration inspired RoboCleanCheck and a longer-term ambition: a verification layer that could help autonomous cleaning systems decide what needs attention next.
What it does
RoboCleanCheck is designed to review photographic evidence across multiple floor zones and cleaning sessions.
The Robo-Clean-Supervising-Agent coordinates five functional modules:
- Robo-Clean-Zone-Audit: organizes photographs by zone and session and identifies missing or ambiguous evidence.
- Robo-Clean-Image-Check: inspects photographs, compares visible conditions, and locates possible remaining debris.
- Robo-Clean-Pattern-Check: identifies comparable problems recurring across distinct sessions.
- Robo-Clean-Action-Plan: proposes prioritized interventions or requests for clearer evidence.
- Robo-Clean-Resolution-Check: uses follow-up photographs to assess whether previously reported problems remain visible.
Robo-Clean-Critic requests a separate visual inspection and challenges unsupported findings. The supervisor limits reinspection and preserves inconclusive outcomes when evidence remains insufficient.
These are logical roles within one UGREEN Skill package. The planned workflow keeps photographs, reports, and history together on NAS and requires local image processing. Analytical conclusions do not require human approval; saving and overwriting reports remain subject to authorization.
How we built it
We developed the workflow through iterative design and review, using ChatGPT and OpenAI Codex alongside the official Uliya Skill Kit and capability catalog.
We authored seven Markdown files and prepared the Round 1 proposal and ZIP package. NAS operations are mapped to the documented file, photo, and knowledge experts.
Official static checking reports zero errors and three warnings concerning values that must be supplied during runtime delegation.
We also conducted exploratory single-image tests using Ollama and Qwen3.5-4B locally. These experiments do not establish UGREEN compatibility. The complete workflow has not yet been executed on a NAS.
Challenges we ran into
A robot having a camera does not guarantee that another application can access its images. Our initial approach therefore uses photographs taken with a phone or fixed camera and made available on NAS.
Visual interpretation also proved difficult: a local model missed small floor details and produced inconsistent findings across repeated runs. This reinforced the need for explicit evidence, critic review, and honest uncertainty.
Accomplishments that we're proud of
We developed a concrete workflow connecting evidence collection, visual assessment, recurring-problem detection, recommendations, and follow-up verification.
We also prepared the submission materials while clearly separating authored instructions, local experiments, and capabilities that still require NAS validation.
What we learned
Reported robot activity does not establish a clean floor. Likewise, an object disappearing from an AI description does not prove that it disappeared from the scene.
Useful verification requires comparable regions, adequate visibility, traceable observations, and a willingness to return an inconclusive result. Agreement between model passes is not proof of correctness.
What's next for Robo Clean Check
If selected, we will validate local image processing and the complete delegation workflow in the assigned UGREEN NAS environment.
Tests will cover partial debris removal, recurring problems followed by resolution checks, and photographs with unsuitable lighting or viewpoints.
Longer term, we want RoboCleanCheck to become a verification component in broader autonomous cleaning systems. Automatic image collection, home-automation connections, and targeted robot actions are future directions requiring separate implementation and validation.
Built With
- openai-codex
- ugos-skill-creation-tool
