Inspiration
"It worked yesterday." I've heard it from others and said it myself more times than I can count. Three experiences stuck with me.
First, I inherited a team project at school. The GitHub repo was messy, the notes were vague, and nobody had written down what changed or why.
Second, I worked on a project that combined a LiDAR (a laser sensor that measures distance) with a camera to find objects and how far away they were. It worked one day and failed the next, and my code hadn't changed. I doubted myself and recalibrated everything. Later I found out that another team sharing the setup had changed something. Nobody did anything wrong. There was just no record.
Third, every week I listed my work for a research presentation, and I often forgot fixes I'd made just days before.
Small physical changes cause big problems. If a sensor is off by just 1°, an object 50 m away shows up about 0.9 m from where it really is. The clues to what changed usually exist, but they're spread across files, photos, voice memos and cameras.
What it does
RoboVault is an AI Skill for the UGREEN AI NAS, the shared storage box that university labs and small R&D teams use. After each test, the team saves its settings files and test report into a folder on the NAS, and records a quick voice note if they touched the hardware. When something breaks, a researcher just asks in plain words: "The robot found objects fine on Tuesday, but today it keeps missing them. What changed?" No commands or technical IDs (like camera IDs, tag IDs or file IDs) needed. RoboVault looks those up itself.
RoboVault compares the good test with the bad one, using the files, photos, voice notes and bench camera already saved on the NAS. Then it writes a short report:
- the most likely cause
- how sure it is
- what it ruled out
- what to test next
The report never blames a person, and RoboVault doesn't save anything unless the user says yes.
How I built it
RoboVault is one SKILL.md workflow, built with the UGOS Skill Creation Tool and Claude. It uses five of the NAS's built-in AI agents. All of the AI runs on the NAS itself, so research data never leaves the lab.
I also tested the problem on real data. I took a public KITTI driving clip and turned the LiDAR 1° to the side, but left its calibration file (the file that says where the sensor is mounted) unchanged. Before the change, every measured distance was within 1 m of the truth. After it, a cyclist 13.4 m away was measured at 19.1 m. At 2°, more than half of the distances were wrong. Yet the calibration file looked exactly the same.
Here's what a RoboVault report on this test would look like:
- Likely cause: physical. The LiDAR mount probably shifted, so a cyclist 13.4 m away was measured at 19.1 m.
- How sure: guess. No file changed, but no photo or voice note confirms the move yet.
- Ruled out: software, settings and test conditions. They're the same in both runs.
- First test: check the LiDAR bracket, then run a quick LiDAR–camera alignment check before recalibrating everything.
Instead of days of guessing, the team knows where to look first.
Challenges I ran into
- UGREEN Skills can't run code. My first plan was for RoboVault to process raw sensor data, like LiDAR scans. But a Skill can't run its own programs. So I changed the plan. RoboVault now works like a detective: it reads the records teams already save (settings files, test reports, photos and voice notes) and pieces together what happened.
- The most important clue is usually missing. When someone bumps a sensor, nobody writes it down, so there's no note to find. Instead, RoboVault works by elimination:
- It checks whether any software or settings file changed.
- If no file changed and the test conditions were the same, but the results got worse, the cause is probably physical, like a moved sensor.
- Then it looks for proof: a photo of the robot, a voice note, or a time when someone was at the bench.
- Cameras can make people uncomfortable. The bench camera is useful because it shows when someone was near the robot. But nobody wants to feel watched. So RoboVault only looks at when a person was at the bench. It never tries to find out who it was.
Accomplishments that I'm proud of
Every claim RoboVault makes points to a specific file or time, so anyone can check it. Reports say what changed and when, never who. And every approved investigation adds to a hardware history, which is the record I wish I'd had back then.
What I learned
- My old tools didn't answer my real question. They showed what the robot recorded. They didn't show what changed on the robot, so finding the cause still took a long time.
- Blame keeps people quiet. When a robot breaks, the fastest fix is often someone saying, "Oh, I moved that sensor yesterday." But people stay quiet when they're afraid of getting in trouble, and asking "Who touched it?" sounds like an accusation. So RoboVault's reports talk about the change ("the LiDAR shifted on Tuesday"), never the person, and nobody gets singled out. When it's about the robot and not a person, people feel safer speaking up, and the team finds the cause faster.
- Trust comes from evidence, not confidence. People trust AI more when it shows how sure it is and why. So RoboVault bases how sure it is on the evidence it finds. If a cause is backed by proof, like a changed file or a voice note, RoboVault says so. If it's only a guess, it says that too, and it lists what it still doesn't know. An answer that only sounds confident is harder to believe.
What's next for RoboVault
Next, I'll test RoboVault on a real iDX6011 Pro and make it even easier and safer for everyday lab use.
