Bedrock-RL pairs Netherite, a deterministic C/CUDA reimplementation of Minecraft, with the verl distributed reinforcement-learning library to train and benchmark vision-language-model agents. Tasks are defined in YAML with swappable reward, renderer, tool, and training-method layers, and the framework generates reproducible synthetic training data from scripted experts while filtering out decisions that would require privileged information. It supports GRPO, supervised fine-tuning, and on-policy distillation, with results evaluated against sealed test splits using paired statistical testing.
