LaelaZorana2024cc-by-4.0

VLA Action Anomaly Eval

A labelled evaluation set pairing real teleoperated robot actions from DROID with realistic faults injected into held-out real actions, designed to measure runtime safety supervisor performance for vision-language-action policies.

Downloads70
Episodes9000

Why This Matters for Physical AI

This dataset enables measurement of runtime safety supervisors for vision-language-action policies by providing real in-distribution actions paired with realistic synthetic faults, addressing the critical need to detect out-of-distribution policy failures in embodied AI systems.

Technical Profile

Modalities
proprioception
Robot Embodiments
Franka Panda
Action Space
end_effector_delta
Environment
lab
Task Types
manipulation
Episodes
9000
Data Format
parquet
Annotation Types
action_labelsreward_labels
License
cc-by-4.0
Part of the DROID family

Access

Need custom proprioception data?

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

Request a Sample Pack

Related Datasets