ENGINEERING NOTES / OCTOBER 2026
Slopebot
A browser-game control experiment that reads a Unity WebGL canvas, estimates a safe corridor, steers, and learns from recorded failures.
01 / THE PROBLEM
Reliable steering starts with reliable perception.
The prototype samples the rendered canvas directly, detects road boundaries and red obstacles, and estimates a lateral corridor. A controller steers toward it while telemetry and death screenshots expose failure modes.
02 / THE SYSTEM
From input to a checked result.
ObserveSample the Unity canvas with preserveDrawingBuffer enabled.
EstimateDetect road geometry, obstacles, and a safe lateral corridor.
ControlApply a PD-style steering policy through browser keyboard input.
EvaluateRecord runs, classify failures, and replay fixtures against a baseline.
Python · Playwright · Canvas perception · Feedback control
03 / EVIDENCE & LIMITS
What has been demonstrated.
Canvas perception → obstacle and road geometry → PD-style steering → game-over detection and restart; telemetry and offline replay support repeatable evaluation.
Perception is heuristic and depends on the 960×540 layout. This is a local research prototype; no general game-playing or reinforcement-learning result is claimed.