NeurIPsMay1234MIT

openpi-interpretability-data

Interpretability artifacts (activations, conceptors, linear steering vectors, sparse autoencoder vectors and checkpoints) extracted from open vision-language-action (VLA) policy models on the LIBERO, MetaWorld, and RoboCasa benchmarks.

Downloads10

Why This Matters for Physical AI

This dataset provides interpretability artifacts from vision-language-action models trained on standard robotics benchmarks, enabling research into understanding and steering learned robotic behaviors through mechanistic interpretability techniques.

Technical Profile

Modalities
rgblanguage
Action Space
language
Environment
simulationlab
Task Types
manipulationpick_and_place
Data Format
PyTorch
Annotation Types
language_instructionsaction_labelsreward_labels
License
MIT
Part of the openpi-interpretability-data family

Access

Need custom rgb data?

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

Request a Sample Pack

Related Datasets