SKU RLE-1502 · Sold by External
Crafter
Product specifications
| SKU | RLE-1502 |
|---|---|
| Data type | RL environments |
| Volume | 22 achievements (evaluation tasks) |
| Size on disk | Not published by source |
| Format | Python package / OpenAI Gym environment; uint8 image observations, procedurally generated worlds |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license (MIT) |
| Quality score | — |
| License | MIT License |
Crafter is a Minecraft-inspired open-world survival game designed as a single-environment benchmark that evaluates a broad spectrum of agent capabilities (exploration, generalization, representation learning, and long-horizon credit assignment). Worlds are procedurally generated 2D grids where the player must forage for food and water, find shelter to sleep, defend against zombies and skeletons, mine materials, and craft a tech tree of tools. The environment follows the OpenAI Gym interface with 64x64x3 RGB image observations and 17 discrete actions. Progress is measured by 22 achievements; the sparse reward is +1 the first time each achievement is unlocked plus +/-0.1 per health point regenerated or lost. Agents get a budget of 1M environment steps and are scored by per-achievement success rates and their geometric-mean 'Crafter score' (human experts reach 50.5%). It ships as the pip package 'crafter' (CrafterReward-v1 and CrafterNoReward-v1 variants) authored by Danijar Hafner and released alongside the paper 'Benchmarking the Spectrum of Agent Capabilities'.