What Is in Movens-5M
The team built the dataset from raw Internet videos. The paper describes two layers: 15 million human motion sequences extracted from the videos, and 5 million humanoid robot motions that were transferred to the robot and found physically executable. According to the project page, the total frame count exceeds 1.08 billion and the median clip lasts 16 seconds.
The motions are grouped into 297 types under 7 families and 16 scene groups. According to the Hugging Face card, each record holds four video views, from the original human video to the robot's tracking rollout in simulation. Written captions for the motion and the video are attached to the record as well.
How the Motions Were Verified in Simulation
According to the paper, the data engine has five stages:
- clip generation
- quality filtering
- semantic annotation
- motion recovery
- physical refinement
In the last stage, human motion is retargeted to the 29-joint motion representation of the Unitree G1. Each retargeted candidate was then replayed in simulation.
A candidate stayed in the dataset if the selected motion tracking policy followed it with sufficient accuracy in simulation. Full validation can evaluate up to four policies (SONIC, Humanoid-GPT, HoloMotion and OmniXtreme), and one success among them is sufficient. A candidate showing a fall or a large tracking error can get a new retargeting attempt before it is rejected. The project page refers to six physical checks. According to the paper, the refinement during retargeting also accounts for joint position and velocity limits, ground contact and foot skating.
Movens-Zero: A Motion Model for the Unitree G1 Humanoid Robot
According to the code repository, Movens-Zero generates whole-body motion for the G1 from text, speech, music or video. According to the project page, a frozen omni-modal context encoder conditions a flow-based action expert in its architecture. The encoder is built on the Qwen2.5-Omni-3B model. According to the paper, the action expert was trained jointly across seven generation tasks and is designed to run in real time, linking motion chunks as instructions or the robot state change.
The team released the model in four sizes:
| Size | Action expert parameters |
|---|---|
| S | 64.2 million |
| M | 187.9 million |
| L | 500.1 million |
| XL | 910.5 million |
According to the paper, results improved as verified data was scaled from 1% to 100% and the action expert from 64 million to 911 million parameters. The Movens-Intent benchmark built for evaluation has a separate track for each input type and 8,192 expert-verified pairs in total.
The paper's appendix shows 30 rollouts on a physical Unitree G1. The authors state that these examples illustrate the range of behaviors rather than estimate a hardware success rate. They list joint-limit compliance in motions generated from music, and additionally contact consistency in motions generated from speech, among the remaining limitations.
Access and License
The dataset can be downloaded from Hugging Face without an access request. The main repository holds a Parquet preview, while the full release is split into tar shards across four separate dataset repositories. The dataset card marks the license as "other" and states explicitly that no commercial-use or redistribution permission should be inferred from that label. Because source videos may contain identifiable people and third-party content, privacy and copyright obligations rest with the user.
The code repository states that license terms for the project code and model weights have not yet been specified. The underlying Qwen2.5-Omni model comes under the Qwen Research License, which permits non-commercial research and evaluation use. Teams planning to use the data or the model should review these terms first.
Unitree G1 Motion Research in Turkey
Movens-5M provides motion data already retargeted to the G1's joint structure. University labs and R&D teams in Turkey can try the data and the model in simulation first and then move to a physical robot. The dataset card also notes that successful tracking in simulation does not guarantee safe transfer to physical hardware. The team's code has separate sections for MuJoCo simulation and for the physical G1. The real-robot setup uses a GPU inference server and a whole-body controller running on the robot.
Robotlar.org is the official Unitree distributor in Turkey and is not among the authors of the Movens work. For the G1's technical specifications, see the Unitree G1 page. We cover motion tracking methods in the G1 motion imitation guide, and real-time motion generation in the MotionBricks guide. For another motion dataset used with the G1, see the BONES-SEED guide. You can ask us for a G1 demo, technical information or a quote for your lab. We work out the robot's configuration with you during the discovery phase.
Sources
- Wendong Bu et al., Movens-Zero: Scaling Humanoid Motion Foundation Models with Large-Scale Human Videos, arXiv:2610.41877 (October 30, 2026)
- Full text of the paper (arXiv HTML)
- Movens-Zero project page
- Movens-5M dataset (Hugging Face)
- Movens-Zero code repository (GitHub)
- Movens-Zero-XL model weights (Hugging Face)
- Unitree G1 official product page

