Skip to content

Releases

OpenCarMind 2.5.1b (beta): ESP32-S3 car board firmware

Choose a tag to compare

@AlveeeRahman AlveeeRahman released this 13 Nov 00:57
735a120

Beta. For the Waveshare ESP32-S3-LCD-1.47B car board only. Do not flash it on the ESP-Mosaico (that board's firmware is OpenCarMind EdgeOS 2.1.0).

OpenCarMind 2.5.1b is a beta of the ESP32-S3 firmware with every change on main since the 2.5.1 release. The parts were tested one by one while they were made, but this image as a whole has not been checked on an S3 board yet. If you want the tested firmware, keep 2.5.1, which stays the latest release and the one the desktop app installs. The board shows "2.5.1b" on page 5.

This release has the firmware only. The desktop app for 2.5.1 works with it.

A new brand, a screen made for the LCD, and a visible power-off.

  • Connects on its own when a car is awkward: on cars where several computers answer at once (engine and transmission, hybrid systems), the logger now picks the engine (7E8, or 18DAF110 on 29-bit CAN), checks that it answers RPM on its own address and polls it directly, so the supported-PID list and every reading come from the engine. With no engine data for a minute it searches by itself: every protocol, each computer on its own address, RPM on each, and keeps the one that works (once per connection; read-only, nothing is written to the car or the adapter). A BOOT double-click starts the same search at any time, or shows the link when it is working. With no adapter in range it lifts the saved-adapter lock for 90 s so a different adapter can be found; only devices that look like OBD adapters are tried. The result is on screen and in the session file (search, poll_ecu).
  • The LED and CPU-monitor tasks keep their stacks in PSRAM, leaving internal RAM for Bluetooth, Wi-Fi and the search; a task that cannot start is now reported instead of failing silently.
  • Page 1 labels the coolant CAR TEMP and page 5 labels the chip's own temperature SYS TEMP.
  • Page 5 MIPS tile says which core is which: SYS (Bluetooth, Wi-Fi clock, display) and AI (car data, the model, the learner). The CPU monitor now releases on a fixed one-second period. The JOBS line shows the job load as SYS/AI in the same order, instead of "core load".
  • Light theme reworked for the LCD: pure white view with tiles drawn by their borders, brand navy text, a light blue accent, warm oranges in place of dark browns, a coral power bar, and vivid section colours on page 5. The Mosaico concept shows its three screens in the dark, light and dark themes.
  • Brand: a new night-drive palette (navy, warm ink, cobalt, amber) across the README, the website, the desktop app and its icons, and a sharper vector logo. Banner chips now read ESP32-S3, Bluetooth LE, TinyML, Desktop app and Active Learning.
  • Screen colours for a backlit LCD: the dark theme is pure black, because any tint in the background glows through the backlight. Text is 4.5:1 or more on every background in both themes. Page 5 gives each of its seven sections its own colour (classifier orange, estimator red), checked for normal and colour-blind vision.
  • Power-off you can see: holding BOOT shows "Turning off" while the drive is saved, then "Off, safe to unplug". The backlight fades and the LED stays dark in deep sleep.
  • Page 1: CPU (both cores, out of 100 %) and estimated power, each as a name, a bar and a value. The connection and MIL line moved to the Faults page. Coolant temperature shows in °F where the units are US.
  • Page 5: instructions per second on each core from the chip's hardware counters, PSRAM used out of total, the chip's own temperature (°C or °F), uptime.
  • Page 3: the pattern line sits above the page dots.
  • New boards start on CarMind R.E.V when the firmware carries it. A board that already chose a model keeps it.
  • Supply chain: CodeQL, a fuzz harness over the input parsers, Dependabot, hash-pinned test dependencies and release provenance.
  • USB door opens only with a code from the desktop tools: HELLO now answers with a random challenge and nothing else (no firmware version, board id or pairing state). The door's code is the HMAC-SHA256 of that challenge keyed with the paired app's token, which only the desktop tools hold; the board then says who it is. The 6-digit code on the board's screen now only pairs a new computer: after it, PAIR is the one command accepted.
  • USB door: unbiased pairing codes. The desktop app explains an expired session and can restart the board.
  • Desktop app: says new boards start on R.E.V, and shows the running model's mark.
  • Report image: ghcr.io/alveeerahman/opencarmind-report serves the SD card's drives as the HTML report (dashboard/Dockerfile).
  • Fix: the switch to the light theme reset the board. Re-colouring every screen created a new LVGL style on every object and part, which emptied LVGL's 64 KB pool and halted in its allocation check until the task watchdog fired 30 s later. The walk now re-colours only what each object set itself, with one style refresh per screen: 27 ms, no allocation. It had been in the firmware since 2.0.1 but only runs once the board knows the time of day.
  • Top bar clock: 1 px more letter spacing, so the digits of the faux-bold time no longer run together.
  • Fixes: the status LED no longer drops a frame when the button and the LED task update it at once ("RMT transaction is busy"). The boot log no longer lists the saved Wi-Fi networks. The model self-test log says plainly when the board differs from the PC simulation.
  • Figures: the seven workflow figures were redrawn with straight connectors and larger text, the three README GIFs re-rendered with the brand fonts, and the screen mock-ups and the Mosaico concept updated to the new colours.
  • Model marks: small variants for figures, tags and GIFs, and figures rendered at 3x, so the marks stay sharp when zoomed.

Flash it

You need esptool 5.6.0 or later.

  1. Connect the board over USB. Hold BOOT while you plug it in if esptool cannot reach it.
  2. Write the merged image: esptool --chip esp32s3 --port <port> write-flash --flash-mode dio --flash-freq 80m --flash-size 16MB 0x0 opencarmind-2.5.1b-firmware-merged.bin
  3. The merged image covers everything from 0x0 to the end of the app, including the settings area, so the board's settings, its pairing with the desktop app and its saved Wi-Fi networks go. The drives on the SD card are not touched. To keep the settings, write only the app at 0x20000: esptool --chip esp32s3 --port <port> write-flash 0x20000 opencarmind-2.5.1b-firmware-app.bin

To go back, flash the 2.5.1 release the same way. Check the download against SHA256SUMS.

OpenCarMind EdgeOS 2.1.0 for ESP-Mosaico (ESP32-S31)

Choose a tag to compare

@AlveeeRahman AlveeeRahman released this 13 Nov 00:09
1053f9d

Updated 13 November 2026: fixed build

The images on this release were replaced on 13 November 2026 with a fixed build from main (bcd05ad). The version name stays OpenCarMind EdgeOS 2.1.0. The release tag still points to 1053f9d, the commit of the first build, and the source of the fixed build is main at bcd05ad. If you downloaded it before, check your file against the checksums below and replace the first build.

Fixed

  • Restart loop in the car. Once a drive had saved its driving-pattern file, every start crashed about 14 seconds later with a stack overflow in the task that talks to the car, and safe mode did not stop it. The pattern file is now read without the large copy, that task needs about half the stack it did, and safe mode turns the pattern code off. Each start also writes one line to obd/resets.csv: the reset reason, how long the board ran, the task that crashed, and the battery voltage, temperature and current.
  • Stopping a drive. Hold End drive for 1 second, or use Settings > LOGGING AND POWER > Stop logging. The board stops the car link, saves and closes the drive files, pauses the battery log and waits until the storage is quiet, then shows "Safe to unplug". It takes 35 seconds at most. If something blocks it, it says "Not safe to unplug yet" with the reason and a Try again button. Logging stays off until you tap Start logging or restart the board. The hold now also works when you lift your finger after 1 second or move it slightly.
  • Turn off. Settings > LOGGING AND POWER > Turn off stops logging the same way, stops the adapter search and darkens the screen once it is safe to unplug. A tap wakes it. The board's POWER button still switches the power, but the firmware cannot see it, so use Turn off first.

Checked on the board. A bench build with a simulated car reproduced the restart loop with the board's own pattern file (5 crashes in a row), and the fixed firmware loaded the same file with no crash. A 13.5-minute simulated drive ran with no crash and 2,836 bytes of the car task's stack never used. Every stop test reached "Safe to unplug" in 1.0 to 1.6 seconds, also while the car kept answering, and no new drive started after it. The fixed release firmware runs on the board with boot health good. Still to check in a real car: a long drive, stopping with your own adapter, and Turn off.

For the Espressif ESP-Mosaico (ESP32-S31) only. For the Waveshare ESP32-S3 car board, keep using the OpenCarMind 2.5.1 release. Do not flash this image on an S3.

OpenCarMind EdgeOS 2.1.0 is the next version of the ESP-Mosaico firmware. It keeps everything in OpenCarMind EdgeOS 2.0.0 and adds three AI services and a Health app. It is not related to Ubiquiti's EdgeOS router software.

Items marked "built, being checked on the board" passed their tests on a computer and run on the board, but have not shown their real result there yet.

The first build of 12 November was built from main (1053f9d). The board checks below ran on build 81017a5, which differs from this image only by one icon table's linkage (a code-scanning fix) and the build stamp. The ESP32-S3 firmware is unchanged.

What is new

Health app

  • A new app right after Board, with three pages you pick with the tabs under the title: Battery, Guard and AI.
  • Each page opens with one status card in plain words: a colour, an icon and a word (Good, Learning, Check soon, Act now), a headline, what it means and what to do. For example "Battery is wearing out", "Holds 76 % of its rated charge, runs shorter on battery", "Plan to replace the cell".
  • The header counts what needs you: "All good", "1 thing to check", "2 things to check". It is amber or red by the worst one.
  • Recent issues: the last five of each page with their time, newest first, and when each one ended.
  • When a battery wear, a battery alarm, a guard block or an adapter question starts, the banner shows "Check soon" or "Act now", the headline and Open Health. A tap opens the right page, a swipe closes it. While you drive, amber issues only go into the list and red ones show for 10 seconds.
  • The Health icon on Home gets an amber or red dot while an issue is active.
  • The details stay below the card for checks: charge, current, power and temperature, the health estimate, the resistance, the anomaly check and its score, the guard's counts, score bars and log, and each AI feature's runs, late runs, worst time against its deadline and work per run against its budget.
  • Read-only. The only button is Clear block.

Battery health from your own charges (built, being checked on the board)

  • A small model estimates the battery's health from the shape of a charge: how much charge goes in between 3.9 V and 4.1 V, where the charge curve has its steepest part, how long the last part of the charge takes and how fast its current falls, the resistance after a load starts, and the temperature.
  • It learns your cell from its first 10 deep charges and keeps that in the board's memory. A deep charge starts below about 3.6 V and ends full. Until then Health says "learning (n of 10 charges)".
  • When a charge from nearly empty to full has been counted, that count is used instead, because it is more exact.
  • The model was trained on lab cells, not on this one. Read its figure as a trend for your cell, not as an exact value.

Battery anomaly detector (built, being checked on the board)

  • Looks at the battery once a minute: the voltage against what the estimator expects, the drift at rest, the current at rest, the heating, the resistance, the load and read glitches.
  • Learns your cell's normal for 24 hours first. The learning is kept, so a restart does not start it again.
  • Only a small autoencoder can raise the alarm. On the test data it gave about one false alarm in 400 hours. Two simpler scores are logged for checks only.
  • An alarm writes a message, and Health names what looks unusual in plain words: it warms up more than usual, loses charge while resting, its current reading looks wrong, its voltage drops more under load, the board draws more power than usual, the connection glitches, or the charge reading drifts. It never switches anything off.

Network guard (built, being checked on the board)

  • Watches what the board's radios already receive: the Bluetooth adverts during a scan, the car link's response times and what a Wi-Fi round sees. The radio code only counts. No address or network name is kept or logged.
  • Learns for two hours after every start and blocks nothing in that time.
  • After that it blocks advert floods and address rotation attacks, Wi-Fi beacon floods and Wi-Fi deauth bursts.
  • A block stops new adapter offers and re-pair requests, listens only to your paired adapter, stops a Wi-Fi round before it sets the clock and keeps the clock, and keeps rows out of the learning.
  • A block never cuts your car link, never sends anything, never pairs and never changes the clock. It lifts 10 minutes after the last suspicious window, or at once with Clear block.
  • A car link that answers much slower than usual gives a note and keeps those rows out of the learning. It never blocks.
  • Settings > Network guard: blocking on or off. On by default. Off means it only logs.

AI services

  • All AI services share one worker below the screen and the models and are admitted by response-time analysis, as in 2.0.0.
  • The admission self-test now runs five times and the fastest run decides, with one more try 30 s later. On one boot of seven in the 2.0.x board check, a single slow run had refused the battery estimator for the whole boot.
  • When deadlines are missed the AI services slow down first. The network guard then checks every fourth window.

Fixes

  • The network guard no longer scores the first Wi-Fi scan after a start. It had nothing to compare with, so every network looked new (290 of 256 right after boot).
  • Health no longer compares a response time with a CPU budget. The AI card says "on time" against the deadline, and the details say "worst time 12.9 ms (limit 1,000 ms)" and "work per run 0.23 ms (budget 2 ms)".
  • The battery current at rest reads "resting, 0 mA" instead of "+0 mA charging". It reads "charging +n mA" or "on battery -n mA" otherwise.
  • The degradation ladder's message fits the event log (it was cut at 55 characters).
  • "CHARGE n mA" on Home and the AI card on Board fit their boxes.

Checked on the board

Build 81017a5 (Health in plain words), flashed through the 1200-baud door with no buttons, every time.

  • Health on the board: the real states (battery learning, guard learning this place, AI on time), and on a debug build that puts a state into the Health view only (never into the battery gauge or the settings): battery good, worn, three anomaly reasons, guard watching, blocking and adapter mismatch, AI slowed, in both themes. The banner's Open Health opened the right page, and a swipe closed it.
  • The first Wi-Fi scan after a start was not scored (16 networks seen, 0 scans scored, Guard shows "no data").
  • A 10-minute soak: boot health good at 68.8 s, the AI services admitted on core 1 (5.00 ms against 1,000 ms), no restart, no USB drop, no missed screen frame deadline, and no missed AI deadline in 1,181 AI runs. The 1200-baud door worked from the new build.
  • Internal RAM stayed above 60 KB. The lowest was 78,003 B, during a Wi-Fi round with the Health app open.

Build 73db232, flashed through the 1200-baud door with no buttons, every time.

  • Boot health good at 68.8 s. All three AI services admitted on core 1: worst-case response 5.00 ms against a 1,000 ms deadline. Self-tests (the fastest of five runs) 229 µs, 19 µs and 58 µs. The first run of each was 2 to 7 times slower, which is why the fastest run now decides.
  • A 10-minute soak: no restart, no USB drop, no missed screen frame deadline, no missed scheduler deadline, and no missed AI deadline in 1,181 AI runs.
  • Inte...
Read more

OpenCarMind 2.5.1: ESP32-S3 car board firmware and desktop app

Choose a tag to compare

@github-actions github-actions released this 05 Nov 04:40

CarMind R.E.V, corrected and harder to break. The model name stays CarMind R.E.V.

  • CarMind R.E.V, corrected classifier head (6e6d698c44c5, replaces f70bdcd3151b): R.E.V no longer calls steady cruising with the throttle open "braking". On two held-out drives of my car, cruising 10.7 % → 87.5 % right and all labelled windows 78.3 % → 95.1 %; on the public KIA test set accuracy 88.9 % → 87.1 %. Only the classifier head changed (126 bytes), so learned fuel, airflow and tank state carry over. See "Looking inside R.E.V with explainable AI" in the README.
  • Outlier-robust fuel learner: Huber weights on a running error scale, pairs with an error over 10 times that scale dropped, labels above 120 L/h ignored (also by the source picker), refits judged with the Huber loss. With 1 % implausible labels on 31 public test cars, fuel R² with the airflow hidden stays 0.867 (2.5.0: 0.696). Learned state from 2.5.0 loads unchanged. REV-STATUS reports rejected=.
  • Checked against meta-learning: ANIL and Meta-SGD (learn2learn) on the same frozen trunk scored 0.778 and 0.197 on clean labels, so the closed-form learner stays.
  • tools/gen_ml_params.py: the learned-state id can name an earlier model through prior_sha256 in export.json, for model changes that keep the trunk and the estimate head.

OpenCarMind 2.5.0: ESP32-S3 car board firmware and desktop app

Choose a tag to compare

@github-actions github-actions released this 05 Nov 02:14
f34e2ba

Two on-board models, chosen in the desktop app: CarMind E.V (fixed, the default) and CarMind R.E.V, the first model that keeps learning in your car.

  • CarMind R.E.V (19,815 parameters, 65,648 bytes, f70bdcd3151b, about 2.9 ms per run): learns your car's fuel use on the board with a fuel read-out that is refit while you drive and goes live only when it beats the current one on held-back rows, plus an airflow constant for cars that lose MAF. Page 5 shows the learned fuel rate. Switch models on the desktop app's Board page while parked; the choice is kept after a restart.
  • Engine torque from physics (both models, shown by R.E.V): fuel energy flow over engine speed minus friction, sized to your engine through the learned airflow constant; uses the car's own torque PIDs (0x62, 0x63) when it reports them. Amber, "idle" when stopped. New CSV columns torque_nm, torque_src.
  • Fuel tank learner (both models): learns the tank size from the fuel level against the fuel counted; shows litres left and range on the details page once it is sure, and suggests fill-ups it sees. New CSV columns tank_left_l, tank_range_km; state in /sdcard/obd/rev_tank.bin.
  • Fill-up calibration (both models): fill-ups typed on the app's new Fuel page give one correction factor for every fuel figure (tank to tank, limited to 0.7 to 1.4). New CSV columns fuel_lph_raw, fuel_cal, trip_fuel_raw_l.
  • Wi-Fi only sets the clock: one HTTPS request to one time server, its certificate checked against trusted roots and its name; the default is www.google.com and the desktop app sets your own (time_host). SNTP is gone, and the board listens on no port. Session JSON clock is now https.
  • Wi-Fi upload to the desktop app is off while it is reworked to avoid head-of-line blocking; import drives over the USB cable. Pairing still protects the USB commands.
  • Page 5: "Learning" job, the time since the model's last answer beside AI ESTIMATES, brake "holding" while stopped, fuel rate in L/h or gal/h with the display units. A model switch clears the last answer, and is refused while a drive is starting or running.
  • Fix: a PID that stopped answering kept its last value, marked valid, until the drive ended. Fields now count as missing after 3 poll cycles (at least 4 s), or 3 rounds of the slow list (at least 10 s), and the CSV leaves them empty.
  • Fix: a MAF reading of 0xFFFF (655.35 g/s, a sensor fault) is ignored everywhere, including by the airflow learner.
  • Fix: R.E.V's learned-state file is read by its own task, not at start-up, which overflowed the main task's stack.
  • Version 2.5.0 for the firmware and the desktop app. Host tests: 524 checks.
  • README: two models, what R.E.V learns and what stays fixed, how the models are made, limitations and future work, larger workflow figures, an ESP-Mosaico concept with R.E.V.

Known limits: R.E.V is in its first drives; torque is an estimate checked on one public car and reads high at a standstill; the tank needs about 10 L of driving before it shows anything; the Wi-Fi clock has not yet been checked on a board.

OpenCarMind 2.4.0: ESP32-S3 car board firmware and desktop app

Choose a tag to compare

@github-actions github-actions released this 04 Nov 04:28
  • CarMind E.V (dfa9af416dfb, 89,152 bytes), a new on-board model that holds up better on cars it has not seen. About 3.5 ms per run. Scores are in docs/MODEL.md.

  • Fixes from a code review, with no change to the model, the scheduler's default behaviour, the OBD polling or the log format:

    • Diesel fuel use from MAF accepts the lean commanded ratios diesels run at (up to 8). It read about 3x high before.
    • A fuel-cut reading of PID 0x44, which is polled slowly, no longer zeroes the fuel rate once the throttle opens again.
    • Fuel and economy start fresh with each drive, and the previous drive's model fuel estimate is no longer borrowed.
    • The amber "estimate" colour is cleared when a value stops being an estimate.
    • Unit captions are redrawn only when the units change.
    • The units event no longer prints the country code when the time zone decided.
    • The 10 s model stall rule applies only when the model runs every second.
    • Host tests: 321 checks.
  • The desktop app shows fuel as unknown, not 0 L, for a drive where the car gave no fuel source.

  • The events screen names the model "ML Model - CarMind Edge". Its hash stays in the serial log and in every .ml.csv row.

  • Repository: removed six screen renders no page used and five desktop design mockups of screens the app does not have. ml/onboard/export.json lists the file details only. .gitignore covers those renders and assistant, scratch and editor files.

  • Architecture proposals: a new issue template. We train what you send on the same data and splits and post the scores.

  • Figures for the inference path updated. The old model's bench summary (ml/onboard/bench_summary.json) is removed.

  • Field check: the author drove the board with the new features (the drive ending when the engine stops, the BOOT power-off, fuel economy, units) and the 10 s model, and reported every feature working.

  • Fuel economy: the first screen's LOAD tile now shows the drive's economy in km/L, or mpg in the US and the UK (LOAD moved to page 3, replacing BARO). The fuel rate comes from PID 0x5E when the car has it, else from MAF with the commanded equivalence ratio (PID 0x44, where 0 means decel fuel cut), else from the on-board model's estimate, shown in amber. New CSV columns fuel_lph_calc, fuel_src, trip_km, trip_fuel_l (main/econ.c, host-tested).

  • Units: metric by default. On the first Wi-Fi connection the access point's country code (802.11d) sets miles and mpg for the US and the UK, once, stored in NVS (main/units.c). A UK time zone is the fallback. Tested on the board with a US access point: it switched to miles and mpg and kept them after a reboot. units in config.json (metric, us, uk, auto) overrides it. Speed, odometer and MIL distance follow the units on screen, and the session JSON records them.

  • Desktop app: fuel used and economy per drive in each drive's units, an average-economy card, and a table of every value the car reported in a drive (columns the car never filled are left out).

  • Scheduler: the model job's stall limit is one 10 s label window instead of three 1 s periods, which the car's own row gaps (up to 3.6 s) could trip.

  • A drive ends when the ECU stops answering for 15 s (before: 11 failed poll cycles, about 2.5 minutes on a real car). The session is saved and closed, the model stops expecting rows, and the board leaves DRIVE, so the power line turns green and the USB door and the Wi-Fi upload work again. An adapter left powered with the ignition off no longer keeps the board in "Not safe: logging". If the ECU answers again without a reboot, a new session starts.

  • Power-off with BOOT: logging stops as soon as the hold reaches 2 s. On release the board waits for the drive to be saved and closed, then sleeps. Holding on to 8 s (forget adapter) resumes logging.

  • DRIVE now means connecting to the adapter or logging rows. Waiting for a quiet ECU counts as parked.