podolinsky

π0.5 LIBERO-10 — prefix features + K=10 action samples (normal + occluded)

Aligned rollouts of the π0.5 policy on LIBERO-10 tasks with K=10 action samples per inference in normal and occluded scene variants, paired with last-layer prefix features from the PaliGemma backbone for representation probing and failure detection.

Downloads18
Episodes500

Why This Matters for Physical AI

This dataset enables research into failure detection and robustness of generative policies through multi-sample action generation and prefix feature extraction, critical for understanding and improving vision-language-action model reliability in manipulation tasks.

Technical Profile

Modalities
videotabularrgblanguage
Action Space
action_chunks
Environment
simulation
Task Types
manipulationpick_and_place
Episodes
500
Data Format
npz
Annotation Types
language_instructionsreward_labels
Part of the LIBERO family

Access

Need custom video data?

Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.

Request a Sample Pack

Related Datasets