bonp_fpv
动态
2026年6月12日 20:57
2026年6月6日 07:33
Specialized Betaflight Skill for Claude LLM: AI-Assisted Tuning and Analysis v0.7.0
👉 https://github.com/SebGalina/betaflight-claude-skill
Hello everyone!
I’ve been working for a while now on a set of scripts implemented as a "Skill" for the Claude LLM.
The core idea is simple: to create a specialized Betaflight assistant that doesn't just answer questions, but actually understands configurations (diff all), analyzes Blackbox logs, and provides hands-on help during the tuning process.
Reliable Troubleshooting For "standard" questions and troubleshooting, it’s already very effective. When asked about Betaflight-related topics, it consults a set of Markdown files that guide its responses based on the official documentation and wiki. This ensures higher reliability, reduces hallucinations, and saves tokens (essentially, we’re "holding its hand" 😄).
It’s great for debugging basic and even many advanced topics. For the truly "nitty-gritty" issues, specialized forums and Discords
Advanced Analysis and Tuning The project gets really interesting when it comes to analysis and tuning.
On the Desktop version (and using Claude Code for those with at least a Pro subscription), I’ve added an MCP server (Model Context Protocol). This acts as a bridge between the LLM and a flight controller (FC) flashed with Betaflight. This means you can interact with your FC directly through the LLM.
Safety first: Safeguards are mandatory—no settings are changed without explicit user validation when a set command is required.
Blackbox Interpretation & Chirp Analysis By using the Python scripts (automated via Claude Code or executed manually), we take it a step further: full interpretation of .bbl Blackbox files, utility scripts, report generation, and more.
The centerpiece is a script called chirp_analysis.py.
Similar to PIDToolBox or what we’re starting to see in upcoming Betaflight releases, it leverages "chirp" logs to generate an incredibly detailed, standalone HTML report. It includes:
Bode plots
Step responses
Phase margins
Crossover frequencies
Stability metrics
Filter analysis
Automated axis diagnostics
Beyond Just Graphs The goal isn’t just to produce another chart. The objective is to generate a report that explains what the data actually means and suggests concrete actions to improve the tune. There are annotations, tooltips, and explanations throughout the report to help users understand the metrics and guide their adjustments.
Another major feature is the incremental approach. With each new tuning pass, the tool can automatically integrate new logs and maintain a complete session history. This allows you to visualize progress, identify regressions, and see the real-world impact of every change.
Live Demo & Scoring If you want to see the output without installing anything, I’ve uploaded a sample report here: 👉 Sample Full Report
I’ve also implemented a scoring system. It’s important to note that this score isn't an absolute measure of "tune quality." It’s a benchmark used within a specific session. A tune with a score of 72 might actually feel better to fly than one that "hits the ceiling" on paper.
Current Limitations The main hurdle right now is that Claude Free Tier users don't have access to Claude Code and cannot run scripts automatically. Claude Desktop also doesn't support automatic execution yet. In these cases, you’ll need to run the scripts manually, but everything is fully documented in the Skill’s README.
2026年5月26日 04:56
I built a Betaflight skill for Claude (AI) — blackbox decode, step-response, diff analysis. Looking for testers & feedback
Hey all,
I've been building an open-source Betaflight skill for Claude (Anthropic's AI assistant) and I'd love to get it in front of people who actually tune their own quads, because that's the only way it gets better.
It's free, Apache-2.0, and not affiliated with the Betaflight project.
The idea: instead of pasting your config into a chat and hoping the model doesn't hallucinate a parameter, the skill ships with real tooling and reference docs so Claude works from your actual artifacts — CLI
dumps, blackbox logs, and plain-language descriptions of how the quad flies.
What it can do today:
- Analyze a diff all / dump — summarizes your build (target, RX/ESC protocol, motor setup, firmware), flags anomalies and deprecated params, suggests changes with rationale (not a blind wall of set commands).
- Decode blackbox logs — it includes a pure-Python port of the official blackbox-log-viewer decoder, so it reads I/P/S frames, gives per-field stats (gyro, motor, eRPM, throttle, accelerometer) and CSV export.
Fully offline.
- Closed-loop step response — setpoint→gyro system identification via Welch cross-spectral method: rise time, overshoot %, settling time, delay, coherence, per-axis diagnosis. Best with a dedicated ID flight
(full-stick steps).
- Diagnose from symptoms — "hot motors after a flight", "wobble on hard yaw", "propwash on descent" → ranked causes + safe quick wins + what to check in the log.
- Setup wizard for a new build, and it can pull up-to-date official presets from betaflight/firmware-presets at runtime.
- Version-aware (4.4 / 4.5 / 2025.12) — flags params that would error on import to newer firmware.
Honest limitations: it's a time-domain tool — no FFT/noise spectra or filter-response plots (use blackbox.betaflight.com or PIDtoolbox for that). It defaults to 2025.12 conventions. It suggests; it never
silently applies anything.
Live FC control (preview): there's a companion MCP server (betaflight-mcp) in the works that talks to your FC over MSP — so Claude can read your PIDs/filters/rates live and, only with explicit confirmation,
write them back and save, no diff copy-pasting. It even has a guided RC-mapping detector (move a stick, it finds the channel) and switch assignment. (Status/link: [à compléter])
Safety — non-negotiable, and baked into the skill: it will never recommend disabling failsafes or arming checks, always warns before motor-direction/mapping changes, always reminds you to test props-off and in
a safe area, and never writes to a live FC without your explicit go-ahead.
How to try it:
- Claude Code: git clone into ~/.claude/skills/betaflight
- claude.ai (web/desktop/mobile): download the attached .zip from the latest release → Settings → Skills → Create skill → upload
- Repo + install details: https://github.com/SebGalina/betaflight-claude-skill
- Latest release: https://github.com/SebGalina/betaflight-claude-skill/releases/latest
What I'm looking for from you:
1. Throw a real diff all or blackbox log at it — does the analysis match what you'd conclude by hand?
2. Where does it give generic or wrong advice? Screenshots of bad answers are gold.
3. Missing parameters, version quirks, build classes (X-class, sub-250, tinywhoop) you'd want covered.
4. Anyone with a bench setup interested in testing the live MCP path when it lands.
Bug reports / ideas: GitHub issues, or just reply here. Tear it apart — that's the point. 🙏
Thanks!
2026年5月21日 06:40
2026年1月20日 12:43
FPVSoundLogger -- On flight demo of my Open-source onboard sound recorder module based on Raspberry Pi Pico RP2024
Soon somes upgrades !
2025年12月17日 05:16
Give Your DJI Footage Real Sound
Hi all !, I’ve pretty much finalized the readme of my little project: it’s called FPVSoundLogger. Here’s the YouTube video showing proof by example:
https://www.youtube.com/watch?v=XYwBHS1GdF4
You’ll find all the details on GitHub (build, how‑to, code, etc.) in the video description. Enjoy, and share if you like it! 😁
2025年12月11日 05:22
Hi everybody, small DIY thing done here ! Because the O4 doesn’t record the quad’s onboard sound — or at least not yet — I tinkered a little thing together. Happy it works perfectly 😁