Event

Research Team Releases Movens-Zero Humanoid Motion Foundation Model

1 reports · 1 sources

The research team released a technical report proposing the Movens-Zero humanoid motion foundation model, which adopts an "understand-generate-execute" architecture and supports real-time closed-loop humanoid motion generation under text, speech, music, and video inputs. The accompanying dataset Movens-5M was built from internet videos through clip generation, quality filtering, semantic annotation, motion recovery, and physics refinement via humanoid retargeting. It contains 15 million human motion sequences and 5 million physically executable humanoid motion sequences, totaling over 10,000 hours, and is aligned with text, RGB video, speech, and music. The team also proposed the Movens-Intent benchmark, covering four categories—text, speech, music, and video—with 8,192 expert-verified condition-motion pairs. The team states that experiments show that when execution validation data is expanded from 1% to 100% and motion expert parameters are increased from 64 million to 911 million, all four modalities achieve consistent improvements, and data quality and cross-modal supervision also improve generation and execution.