Factoriax - A GPU-Accelerated Factory Building Simulator In The Style Of Factorio
Poster A: Monday -- 11:00 - 12:30
Mickey Beurskens, Tristan Tomilin, Thiago D. Simão
Keywords: reinforcement learning, benchmarks, long horizon, curriculum learning, factory planning, JAX, GPU accelerated
We introduce Factoriax, a JAX-native simulator for integrated factory building.
Easy-rocket is its first benchmark, a 16-by-16 grid scenario where laying out machines, routing items between them, balancing production, and operating the factory are coupled into a single task. The preview comprises the simulator, the benchmark, and a scripted baseline that completes the scenario.
A 1-billion-step PPO training run on easy-rocket completes in about 8 minutes on a single NVIDIA A100 and about 84 minutes on a laptop GPU. Across ten seeds at 10 billion environment steps each, a strong PPO baseline saturates at 17.7 of 25 achievements. It learns the placement-rewarded curriculum but stops at the configuration wall, the spatial relations between placements that are essential to build factory lines.
We bring this preview to EWRL ahead of a full release hoping for input from the community to steer future development.