MotionMillion: 2,000 Hours of Annotated Human Motion Data

InternRobotics released MotionMillion in 2025, providing over 2,000 hours of human motion captured from web-scale video data. The collection spans diverse activities including Fitness (26.4%), Martial Arts (23.7%), and Dance (14.9%), featuring more than 20 detailed semantic annotations per sequence.

Use Cases

• Training humanoid locomotion models using the Movement Styles and Body Characteristics labels
• Emotion-driven animation synthesis leveraging the Emotions annotation field
• Action recognition for fitness and martial arts using the category-specific motion sequences

Strengths

• Over 2,000 hours of motion data
• 20+ detailed annotation types including emotions and environments
• Diverse activity coverage with 26.4% fitness and 23.7% martial arts

Limitations

• Potential label noise inherent in web-scale video extraction
• Small representation of non-human motion at only 2.9%
• Lack of documentation regarding specific coordinate systems or skeleton formats

Provenance

Source

InternRobotics

Collection Method

Extracted from web-scale human video data

Freshness

Last updated November 2025.