CASE STUDY 02
Solves Tower of Saviors orb boards in a browser or on Android, then executes the drag.
Solves Tower of Saviors orb boards in a browser or on Android, then executes the drag.
Choosing and dragging an orb route by hand was the problem I wanted to automate.
Two targets share one solver. Only how the board is read and how the drag is performed change between them.
Takes the orb grid straight from the simulator's canvas pixels.
Classifies screenshot patches of the live game into orb types, and tracks sealed, shielded, cursed and blocked cells.
Beam search over drag paths, each scored by an exact combo, gravity and cascade simulation.
The Chrome extension performs the route in the simulator.
ADB and MaaTouch perform the route on the phone, after dry-run and stale-board checks.
CHAPTER 1
CHAPTER 2
CHAPTER 3
Drag paths are scored by simulating the whole resolution — combos, gravity and cascades — not by matching only the first clear. Parallel beam search explores the routes, and constructive planners take over when the beam comes up empty.
The same solving core drives a Chrome simulator and the real Android game. Reading the board differs — canvas pixels in the browser, classified screenshot patches on the phone — and so does the drag, but the decision is shared.
The phone runner composes stage rules and card effects instead of hardcoding a route: clear-all demands, required or forbidden rune types, pinned endpoints, touch conversion, and rearrange mode with several physical drags before the board resolves.
No build step and no package dependencies. A deterministic regression suite covers the solver.
Runs the solver against the browser simulator, reading its canvas and performing the drag.
Parallel search over drag paths, scored by full cascade simulation.