Bedrock-RL brings reproducible reinforcement learning to Minecraft VLM agents

Hugging Face community contributors have introduced Bedrock-RL, an open framework for training and benchmarking vision-language models (VLMs) inside Minecraft. Its core goal is to make embodied-agent research more reproducible: instead of relying on slow, unpredictable game clients and human gameplay videos, developers can run controlled, parallelized simulations with verifiable outcomes.

A modular training stack

Bedrock-RL combines Netherite, a deterministic C/CUDA Minecraft simulation, with the distributed reinforcement-learning library verl. Experiments are configured in YAML, including the world setup, task instructions, reward logic, model, tools, observation view, and training method. Developers can replace individual components without rewriting the rest of the pipeline.

Supported approaches include several policy-optimization methods, supervised fine-tuning, self-distillation, and on-policy distillation. The framework also supports different renderers, including semantic and procedural views, as well as the official Minecraft view.

The project is designed for synthetic data generation as well as training. Scripted agents can produce trajectories across isolated episodes, while the system records seeds, state hashes, and provenance. A verifier filters for successful examples, and safeguards prevent training data from depending on information unavailable to the model, such as a complete map when the agent only receives images.

Why it matters

For AI tool builders, the main value is a repeatable testbed where changes to rewards, tools, context windows, or models can be evaluated against the same environment and verifier. In a small example, GRPO training improved Qwen3-VL 2B’s three-attempt success rate on an item-selection task, although the authors present it as a pipeline check rather than a benchmark result.

Source: Hugging Face Blog