MK7 Bot

Status: compiles against stand-in headers and the geometry/manager logic is tested on a PC, but nothing has run on real hardware. Expect to fix small API mismatches in your CTRPF version.

What's here

• plugin.cpp: the bot (3GX plugin). Learns the track by trailing the lead CPU kart on lap 1, then drives the learned line alone. Self-tunes steering/braking per lap, brakes for sharp corners, reverses out when stuck, uses items by situation, draws the planned line on the bottom screen.
• manager/main.c: the 3DS app. A = inject, B = eject, X = forget learned tracks.
• trainer.py: optional Citra training for a neural net (Citra I/O, env, PPO, export in one file).

Step 0: fill in the config block at the top of plugin.cpp (required)

Kart list address, struct stride or pointer mode, offsets for x/z/yaw/speed/rank/item, race flag, course id. Also item IDs in SetupItems(). Find them via CTGP-7 source or Citra's debugger. Mirror the same kart addresses in the ADDR dict in trainer.py if you use the training path.

Build and install (no eShop needed)

1. Get custom firmware: follow 3ds.guide (Luma3DS). Back up NAND when told to.
2. Plugin: devkitPro + CTRPluginFramework template, build plugin.cpp to MK7Bot.3gx. Put it at SD:/luma/plugins/<MK7 title id>/MK7Bot.3gx.off (note the .off). Check your game's title ID folder on a title database; it differs by region.
3. App: devkitPro libctru, build manager/main.c using the 3ds-examples application template. Run it from the Homebrew Launcher, or make a CIA and install it with FBI to get a HOME menu icon.
4. In the manager press A to inject, then launch Mario Kart 7. Press B to eject. Closing the game always removes the bot from memory.

Using it

Open the plugin menu in game (Select by default), turn on "Bot on/off" in a Vs race with CPUs.

• New track: it trails the lead CPU for lap 1 (status "LEARNING LINE"), saves the line, then drives alone.
• Known track: it drives immediately and keeps tuning ("best cost" on the bottom screen).
• "Forget this track" wipes that course's line and tuning.

Optional neural net (Citra)

python trainer.py bin2json line_XXXXXXXX.bin, then train, then export; copy policy.bin to SD:/MK7Bot/ and enable "Neural policy" in the plugin.

Known limits

• Drifting and mini-turbos are not implemented yet; items use rules, not learning.
• D-pad steering with pulse-width modulation is the default; set HAVE_CIRCLEPAD_INJECT if your CTRPF supports analog.
• Item/aim button mappings (L for item) need checking.
• Online play is gone anyway; this is for offline CPU races only.