DatasetFeatured
LIBERO
Simulation benchmark for lifelong learning and VLA eval with many manipulation tasks.
Dataset
Meta-World
Classic multi-task manipulation sim benchmark (50 tasks) for meta and multi-task RL.
Overview
Meta-World is the classic 50-task manipulation benchmark for multi-task and meta-RL in simulation. Despite its age, it remains a common sanity check for multi-task algorithms because tasks are standardized and compute requirements are modest.
It is less aligned with modern language-conditioned VLAs, but still useful for algorithm papers focused on multi-task RL structure.
Who it is for
Multi-task / meta-RL algorithm researchers.
Key highlights
- 50 standardized manip tasks
- Meta-RL heritage
- Low compute barrier
When to use
- Multi-task RL algorithm development
- Teaching multi-task RL
When not to use
- Language-conditioned VLA SOTA claims alone
Getting started
- 1Install Meta-World
- 2Train multi-task SAC/PPO
- 3Report MT10/MT50 style metrics
Papers & reading
How it compares
DatasetFeatured
RoboCasa
Large kitchen sim suite with generative scenes and imitation/RL data.
Caveats & pitfalls
- Visual and language richness lag modern kitchen/home suites.
Content reviewed 2026-07-23