SKU RLE-1507 · Sold by External
Overcooked-AI
Product specifications
| SKU | RLE-1507 |
|---|---|
| Data type | RL environments |
| Volume | 49 layout files (kitchen configurations; 5 are the canonical benchmark layouts) |
| Size on disk | Not published by source |
| Format | Python simulator (installable overcooked_ai_py package with Gym/PettingZoo-style envs) plus .layout config files (Python-dict text) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license |
| Quality score | — |
| License | MIT License |
Overcooked-AI is a fully-cooperative two-player gridworld benchmark (inspired by the video game Overcooked) built by UC Berkeley's Center for Human-Compatible AI to study human-AI coordination. Two agents move around a shared kitchen, pick up onions/tomatoes/dishes, place ingredients into pots, wait for soups to cook, and deliver finished soups to serving counters for reward. Each agent has 6 discrete actions (four movement directions, stay, and interact), and the environment exposes both a lossless boolean-mask tensor encoding (for CNNs) and a hand-crafted featurized state vector. The repository ships 49 .layout kitchen-configuration files, including the five canonical coordination layouts studied in the paper (cramped_room, asymmetric_advantages, coordination_ring, forced_coordination, counter_circuit). Reward is a shared sparse signal of +20 per delivered soup, with optional dense shaping rewards for intermediate actions. The code is released under the MIT license by the Center for Human-Compatible AI.