Constrained Creativity — ML, game development, and generative systems.
Most of what's here starts from the same question: what happens when you give a learning system a small, strict set of rules and let it find the edges? Some of it turns into games. Some turns into the tooling needed to build the games.
A two-player ring-out fight. Push beats dodge, dodge beats block, block beats push — rock-paper-scissors with a stamina meter and commitment frames, so the whole game is reading what your opponent is about to commit to. The opponents are PPO agents trained to do that reading.
The devlog is an honest account of getting it wrong: a reward economy that quietly killed two-thirds of the move set, bots that couldn't aim, and why self-play was the wrong call.
A Rush Hour puzzle engine and state-space visualizer in Unity. The 6×6 board —
up to 16 cars with lengths, orientations and positions — compresses into a
single ulong, so state transitions are zero-allocation bitwise ops and BFS
prunes through a transposition table. Enumerates the full reachable state space
(~100k states) in under 5ms on one thread.
A framework for building RL environments as batched tensor simulations, with correctness enforced by differential testing. Every environment gets two implementations from one spec — a readable single-instance reference in plain Python, and a batched PyTorch version where branching becomes masking. Both are written from the spec, never from each other, and correctness means bit-identical behavior under the validation battery.
High-resolution images into constrained low-pixel sprites (16×16, 32×32, 48×48). Deterministic path first — area-average downscaling, OKLab palette snapping, hard alpha matting, despeckling, no model required — with an optional learned restoration path for degraded or AI-generated inputs. Ships as a PySide6 desktop app and a CLI.
An elevator control sandbox: train a dispatcher with Unity ML-Agents, run it against heuristic baselines, and watch the two argue on a live dashboard.
orbitope.github.io · orbitopegames@gmail.com