phi-monster2026apache-2.0

Deconfounded goal set (LIBERO)

600 episodes with objects displaced 6-13 cm and fixed instructions, designed to test reading current object positions rather than memorized ones. Addresses position-memorization bias, improving performance from 31.2% to 83.3% on displaced referents.

Downloads33
Episodes600
Hours0.8309

Technical Profile

Modalities
rgblanguage
Action Space
7-DoF: action
Environment
simulation
Task Types
manipulationgrasping
Episodes
600
Total Hours
0.8309
Data Format
LeRobot
License
apache-2.0
Part of the Deconfounded goal set (LIBERO) 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