wendell0218other

OmniHM-5M

Large-scale, multimodal, execution-verified humanoid motion dataset built from Internet videos containing 5 million physically executable humanoid-motion sequences and 15 million human-motion sequences totaling over 10,000 hours. The dataset aligns motion with text, RGB video, speech, and music across 297 motion types and 16 scene groups.

Downloads22
Episodes5000000
Hours10000

Why This Matters for Physical AI

OmniHM-5M provides a massive corpus of physically executable humanoid motions with multimodal alignments (vision, language, audio) that enables training of embodied AI models for motion understanding, generation, and control in humanoid robots.

Technical Profile

Modalities
rgblanguageaudio
Robot Embodiments
humanoid
Environment
internet_videos
Task Types
motion_generationmotion_retargetinghumanoid_motion
Episodes
5000000
Total Hours
10000
Data Format
parquet
Annotation Types
language_instructionsmotion_captionsvideo_captionsmotion_type_labelsobject_labelshuman_demographicshuman_mood
License
other
Part of the OmniHM-5M family

Access

Need custom rgb data?

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

Request a Sample Pack

Related Datasets