DatasetFeatured
RoboCasa
Large kitchen sim suite with generative scenes and imitation/RL data.
DatasetFeatured
CALVIN
Language-conditioned long-horizon manipulation sim benchmark emphasizing compositional generalization.
Why: Classic language-conditioned long-horizon benchmark
Overview
CALVIN is a language-conditioned long-horizon manipulation benchmark in simulation, emphasizing compositional generalization across instructions and environments. It stresses multi-step behaviors more than single short skills.
Use CALVIN when your method claims long-horizon language following, not only single-step pick-and-place accuracy.
Who it is for
Long-horizon language-conditioned policy researchers.
Key highlights
- Long-horizon language tasks
- Compositional generalization focus
- Established sim benchmark
When to use
- Multi-step instruction following eval
When not to use
- Only short-horizon single skills
Getting started
- 1Install CALVIN
- 2Run sequence evaluation protocols
- 3Report full multi-step success
Papers & reading
How it compares
Caveats & pitfalls
- Long-horizon metrics are strict — ablate carefully.
Content reviewed 2026-07-23