Franka Kitchen

A MuJoCo kitchen scene where a Franka arm completes sub-tasks in sequence, such as opening a microwave and moving a kettle.

Franka Kitchen comes from the 2019 CoRL paper Relay Policy Learning (Gupta, Levine, Hausman, and colleagues), which built a kitchen in MuJoCo containing a 9-degree-of-freedom Franka arm (7 arm joints plus 2 fingers) that can operate a microwave door, a kettle, a light switch, a sliding cabinet door, a hinged cabinet door, and stove knobs. Each episode has to complete several specified sub-tasks, earning 1 point per completion, making it a sparse-reward, long-horizon, multi-task setting. It was later incorporated into the D4RL offline reinforcement-learning benchmark, providing complete, partial, and mixed demonstration data, and is now maintained by the Farama Foundation's Gymnasium-Robotics; it's commonly used to test whether offline reinforcement learning, imitation learning, and hierarchical policies can chain multiple skills together.

ExampleIn D4RL's kitchen-complete data, every demonstration completes the same 4 sub-tasks in order — opening the microwave, moving the kettle, flipping the light switch, and pushing open the sliding cabinet door; the mixed data instead completes these 4 sub-tasks out of order and incompletely, testing whether an algorithm can stitch fragments together.

Also called

FrankaKitchen, D4RL Kitchen

Related

D4RL · MuJoCo (Multi-Joint dynamics with Contact) · Long-horizon Task · Offline Reinforcement Learning · Sparse Reward · Franka Emika Panda / Franka Research 3

Sources

Franka Kitchen - Gymnasium-Robotics Documentation
Relay Policy Learning (arXiv 1910.11956)
Minari: D4RL Kitchen datasets

As of

2026-09