CLARU
phi-monsterapache-2.0

Deconfounded compositional set (RoboCasa)

A dataset of 504 episodes featuring attribute conjunction tasks (size × colour) with single-attribute foils for compositional grounding in robotic manipulation. Part of the Galahad release, designed to test compositional understanding where each foil matches the target on exactly one attribute.

Downloads3
Episodes504

Why This Matters for Physical AI

This dataset addresses compositional grounding in robotic manipulation, enabling models to learn how individual attributes combine to identify target objects, which is crucial for generalizable physical understanding.

Technical Profile

Task Types
manipulationgrounding
Episodes
504
Data Format
LeRobot
Annotation Types
language_instructionsattribute_labels
License
apache-2.0
Part of the RoboCasa family

Access

View on HuggingFaceRead Paper

Need custom physical AI data?

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

Request a Sample Pack

Related Datasets

RoboCasa Simulation Assets

3D object models and textures for RoboCasa simulation environment, including Objaverse, Sketchfab, and Lightwheel objects along with environment and AI-generated textures.

375 downloadsApr 2026apache-2.0

RoboCasa Simulation Assets

3D object models and textures for RoboCasa simulation environment, including assets from Objaverse, Sketchfab, Lightwheel, and AI-generated textures.

214 downloadsApr 2026apache-2.0

Deconfounded relation set (RoboCasa)

624 episodes from RoboCasa with randomized anchor and target slots to study spatial relations, where relation properties are deconfounded from position. Part of the Galahad release.

3 downloadsJul 2026apache-2.0

Deconfounded ordinal set (RoboCasa)

A dataset of 504 episodes from RoboCasa focused on ordinal reasoning tasks with randomized spacing and slots, where rank words rather than absolute positions predict targets. Part of the Galahad release designed to deconfound ordinal understanding from absolute position encoding.

3 downloadsJul 2026apache-2.0

Deconfounded negation set (RoboCasa)

A dataset of 517 episodes from RoboCasa featuring negation-grounded language instructions where positive and negated instructions (e.g., 'the apple' vs 'not the apple') select different objects, designed to deconfound negation from keyword matching.

rgblanguage
3 downloadsJul 2026apache-2.0