An AI that plays WeChat Jump, then dies on purpose

What it does This project automates the WeChat Jump Game by locating the character and the next platform using either classical computer vision or a coarse-to-fine CNN. It calculates how long to hold the jump—via a configurable sensitivity parameter—and issues commands to the phone over USB. The loop repeats, letting the AI rack up jumps until it misses or you stop it.

The interesting bit The project treats a casual mobile game as a localization problem, using a coarse-to-fine CNN to pinpoint landing zones. The author also has a sense of humor about the result: the system reached 859 jumps in about 1.5 hours, at which point the author simply chose to let it “go die” rather than chase an endless high score.

Key highlights

• Ships with two approaches: a classical CV pipeline (play.py) and a CNN-based coarse-to-fine model (nn_play.py)
• The neural network path supports both Android and iOS; the classical fast-search path is currently Android-only
• Achieved 859 consecutive jumps before the run was manually ended
• Provides pre-trained models, training logs, and datasets if you want to retrain the localizer yourself

Caveats

• The classical CV fast-search path is Android-only, and the README warns that play.py may fail on some phones
• iOS support requires a Mac and a chain of dependencies the README outsources to a third-party forum post
• Training your own model from scratch is explicitly not recommended without a GPU

Verdict Worth a look if you are studying mobile game automation or CNN-based localization. Skip it if you need turnkey iOS support without wrangling Mac-specific dependencies.