✦
Summary
PlayTrain turns single-file JavaScript games into environments for reinforcement learning. Researchers can use language models to generate or modify a game, play it in a browser, and train pixel-based agents on the same underlying implementation, making it easier to create and study new interactive tasks.
Key Features
An environment can be implemented as one JavaScript file.
Browser play uses the same games used for agent training.
The execution pipeline exposes games through a standard gym environment.
The public examples compare IMPALA and PPO agents.
Language-model-assisted edits can change dynamics, procedural logic, and test settings.
The demonstrations train agents end-to-end from visual observations.
Recorded seeds and action sequences allow replay of human and agent runs.
The evaluated suite exceeds one million agent decisions per second on one GPU node.
About PlayTrain
The framework exposes JavaScript games through a standard gym-style interface and an efficient execution pipeline. Its demonstrations compare IMPALA and PPO training on Atari- and ProcGen-style games, with procedural modifications, repeatable seeds, and human play data supporting controlled comparisons.
PlayTrain is useful for rapidly varying game mechanics, building new evaluation settings, and comparing human and agent behavior. The project reports more than one million agent decisions per second on a single GPU node for its evaluated suite. Actual speed depends on game complexity and training configuration.