mahgoobi

scoring_data

A branch-and-rollback search dataset containing 650 searches across 13 tasks with 24,270 nodes, where multiple action chunks are proposed from each state and evaluated for their outcomes, including counterfactual rollbacks of abandoned branches.

Downloads0
Episodes650

Why This Matters for Physical AI

This dataset enables training of action scorers by providing counterfactual trajectory data showing what would have happened with different action choices, critical for learning robust action selection in robotic manipulation.

Technical Profile

Modalities
rgbstate
Action Space
action_chunks
Environment
simulation
Task Types
manipulationpick_and_placeobject_placement
Episodes
650
Data Format
parquet
Annotation Types
action_labelsreward_labels
Part of the RoboPRO 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