Bedrock-RL makes Minecraft repeatable enough to train vision-language agents

Hugging Face published an August 19 article by Old Dominion University graduate researcher Michael Evans introducing Bedrock-RL, an open-source framework that makes Minecraft experiments repeatable for training vision-language agents by coupling a deterministic C/CUDA reimplementation of Minecraft 1.11.2 (Netherite) with ByteDance's distributed reinforcement-learning framework verl. The framework, hosted on GitHub under an MIT license, supports up to 7,200 lockstep worlds on one GPU and provides fixed worlds, verifiers, and traceable training data across models and methods.

Bedrock-RL makes Minecraft repeatable enough to train vision-language agents The open-source framework keeps worlds and rewards fixed while researchers swap models and trainers, aiming to make Minecraft VLM experiments reproducible. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Hugging Face https://huggingface.co/blog/Michael-E/bedrock-rl Why it matters Bedrock-RL turns Minecraft agent work into a more controlled experiment, giving researchers reproducible worlds, fixed verifiers and traceable training data across models and methods. Michael Evans https://mevansci.github.io/?ref=runtimewire , a graduate researcher at Old Dominion University, detailed Bedrock-RL in an August 19 Hugging Face community article https://huggingface.co/blog/Michael-E/bedrock-rl?ref=runtimewire , laying out an open-source framework designed to make Minecraft experiments repeatable enough for reinforcement learning. The Bedrock-RL repository https://github.com/michaelbf16/bedrock-rl?ref=runtimewire , hosted under Evans' GitHub account and published under an MIT license, packages a deterministic Minecraft simulation, distributed training infrastructure and task-verification system into one stack for vision-language models. The Hugging Face article credits several community contributors alongside Evans. Evans arrived at the problem from a research background where reproducibility carries consequences beyond a leaderboard. According to his curriculum vitae https://mevansci.github.io/pdfs/cv.pdf?ref=runtimewire , he is pursuing a master's degree in electrical and computer engineering and studies multimodal medical imaging at Old Dominion's Vision Lab. His work has included pipelines for Alzheimer's disease research, brain-tumor segmentation and prediction of regional brain glucose hypometabolism. He previously worked on scientific-claim verification with large language models and autonomous-vehicle simulation. That record helps explain Bedrock-RL's defining concern: a model result should be traceable to the world, observation and decision sequence that produced it. Repeatability is the product Minecraft has long appealed to AI researchers because it combines visual perception, navigation, tool use, planning and objectives that can stretch across hundreds or thousands of actions. The standard Java client creates an awkward training substrate, however. Bedrock-RL's authors describe it as slow, nondeterministic and difficult to run across many parallel rollouts. Bedrock-RL couples Netherite https://github.com/Infatoshi/netherite?ref=runtimewire , a deterministic C/CUDA reimplementation of Minecraft 1.11.2, with verl https://github.com/verl-project/verl?ref=runtimewire , the distributed reinforcement-learning framework initiated by ByteDance's Seed team. Netherite's repository claims support for up to 7,200 lockstep worlds on one GPU. Bedrock-RL uses that engine as the controlled environment beneath its training and evaluation tools. Researchers describe a world, instruction, action budget and reward in YAML. Bedrock-RL can then produce separate training, development and test episodes using disjoint seeds. Each saved trajectory carries its world seed, decision seed, snapshot hash and provenance, giving researchers the inputs needed to replay an episode rather than reconstruct it from an incomplete log. The architecture divides an experiment into replaceable layers for the task, visual representation, tools, context policy, teacher guidance, data pipeline, model and trainer. A researcher can change the renderer or learning method while preserving the underlying task and verifier. Supported methods include GRPO, RLOO, REINFORCE++, ReMax, supervised fine-tuning and several forms of distillation. That separation matters because agent benchmarks can become moving targets when changes to the model also change the environment wrapper, prompts, rewards or evaluation harness. Bedrock-RL keeps the verifier fixed across training methods. Reward checks read the live engine state, so success can depend on whether the agent actually selected an item, collected a resource or reached a structure, rather than whether its generated text claimed that it had. Evans and the contributors also built a guardrail for synthetic demonstrations. Scripted expert policies may see a complete voxel map, while a vision-language model receives frames. Bedrock-RL records whether each expert decision was grounded in information available to the model and rejects privileged actions by default. A route based on unseen underground iron, for example, should not become a behavior-cloning target for an agent that could not observe the ore. A small test with useful restraint The included training example asks Qwen3-VL 2B to select an iron pickaxe from nine randomized hotbar items. Bedrock-RL trains the unmodified base model for 10 GRPO steps and evaluates each checkpoint against 288 frozen development prompts. At temperature zero, the base model succeeded on 27 of 288 prompts, or 9.4%. The step-10 checkpoint succeeded on 39, or 13.5%. The reported paired exact p-value was 0.169, leaving that single-attempt gain statistically inconclusive. With three stochastic attempts per prompt, pass@3 increased from 31 of 288, or 10.8%, to 64 of 288, or 22.2%. Bedrock-RL reports 37 paired improvements, four regressions and a p-value of 1.03 x 10^-7 for that result. The authors present the run as an end-to-end pipeline check. They kept the sealed test split closed because no development checkpoint met their selection standard. That decision is more informative than squeezing a benchmark claim from 10 training steps: Bedrock-RL is trying to make evaluation discipline part of the framework rather than a convention researchers are expected to remember. The example establishes that the machinery runs rather than showing broad Minecraft competence. Selecting a hotbar item is far removed from long-horizon navigation, crafting, combat or construction. Bedrock-RL earns attention because those harder tasks can be expressed within the same controlled system. Minecraft gets another research stack Bedrock-RL enters a field with a long history. Microsoft's Project Malmo https://www.microsoft.com/en-us/research/project/project-malmo/?ref=runtimewire provided a Minecraft-based experimentation platform for reinforcement learning and other AI research. MineDojo https://github.com/MineDojo/MineDojo?ref=runtimewire assembled thousands of tasks and a large knowledge base drawn from Minecraft material. Voyager https://voyager.minedojo.org/?ref=runtimewire used GPT-4 /models/openai/gpt-4 , an automatic curriculum and a growing library of executable skills to create a continually exploring agent. Craftax https://craftaxenv.github.io/?ref=runtimewire pursued faster open-ended reinforcement learning in a JAX environment inspired by Minecraft. Bedrock-RL's contribution is narrower and closer to the training loop. It gives researchers a deterministic simulation, seed-controlled data generation and an unchanged verifier for comparing VLM policies and learning methods. The design treats Minecraft as infrastructure for controlled experiments rather than a demo environment where a successful run becomes the result. For Evans, the project also marks a substantial expansion from medical imaging and scientific verification into embodied-agent engineering. His prior work focused on extracting defensible conclusions from noisy scientific data. Bedrock-RL applies the same instinct to an unruly virtual world: record the conditions, separate development from testing and make every claimed success replayable. The framework still has to prove itself on meaningful long-horizon tasks and across research groups outside its contributor base. Its strongest early argument is already visible in the code. Bedrock-RL makes the unglamorous parts of agent research - seeds, verifiers, provenance and rejected trajectories - first-class components. Those are usually the parts that determine whether a result survives contact with another lab.