From Human Motion Capture to a Unitree G1: Retargeting Real Motion — Including Object Manipulation
- Aug 13
- 4 min read
There's a big difference between capturing human motion and producing motion data a robot can actually use. In this post we'll show the full path — from a performer in our studio to a Unitree G1 humanoid picking up and handling objects — and explain why object manipulation is the part that separates usable robot training data from a good-looking render.
The path: from performer to humanoid
It starts with real people. In our studio, performers wearing optical motion capture markers move through specific tasks — walking, reaching, lifting, and handling objects — captured in full body with finger-level detail.
That raw motion is then cleaned, retargeted onto the target humanoid skeleton, and — in the clip above — applied to Unitree G1 humanoids in simulation. The same human motion, now driving the robot. This is exactly how human motion capture becomes Unitree G1 training data for imitation learning and whole-body control.
Why object manipulation is the hard part
Locomotion — walking, turning, balancing — is relatively well covered in existing motion datasets. Object manipulation is not. The moment a robot has to pick something up, the data has to get everything right: the approach, the hand contact, the grip, and how the object's weight changes the body's motion.
If the contact or the object interaction is sloppy in the capture, it breaks the instant the motion goes onto a robot. That's why we verify object interaction carefully — so the manipulation holds up when a humanoid actually performs it. It's the difference between a clip that looks nice and data a robot can learn from.
Delivered pipeline-ready
Capturing and retargeting is only useful if the data drops cleanly into a learning pipeline. We deliver in the formats robotics and embodied-AI teams work in:
SOMA BVH — clean skeletal motion, ready for retargeting
NPZ — array data for ML pipelines
Retargeted skeletons — including Unitree G1, or a custom skeleton on request
Metadata JSON — action labels, performer data, object data, capture parameters
Synchronized reference video
We built our own production pipeline and custom tools to generate this reliably at scale, and we extend the tooling whenever a team needs a specific format, skeleton, or data structure. We don't just capture the motion — we make sure it's usable on your robot.
Custom capture for your robot and your tasks
Most of our work is custom. Every robot platform and every training objective is different, so we design and capture datasets around your specific tasks and morphology — locomotion, manipulation, object interaction, multi-agent, and more — retargeted to your humanoid and delivered to spec.
Try a free sample
We've published a free 5-hour production sample on Hugging Face so robotics and AI teams can evaluate our capture quality, structure, metadata, and formats before starting a conversation.
Frequently asked questions
How do you retarget human motion to a Unitree G1? Human motion is captured with optical mocap, cleaned, and mapped from the performer's skeleton onto the Unitree G1's skeleton, so the robot reproduces the human's movement — including object manipulation. It's delivered in formats like SOMA BVH, NPZ, and retargeted skeletons with metadata.
Where can I get object manipulation data for humanoid robots? From a motion
capture studio that captures and verifies object interaction. Apple Arts Studios produces custom human motion datasets — including object manipulation — retargeted to humanoid platforms like the Unitree G1 and delivered pipeline-ready.
What formats is the motion data delivered in? SOMA BVH, NPZ, retargeted/custom skeletons, and metadata JSON, plus synchronized reference video. Other formats (BVH, FBX, C3D, CSV) are available on request.
Is the motion applied to a real robot or in simulation? In the video and images above, the retargeted motion is shown on Unitree G1 humanoids in simulation. The same data is built to drive physical humanoid platforms through a robot learning pipeline.
Can you capture custom manipulation tasks? Yes. We design and capture datasets around a specific robot and a specific task list, including custom object-manipulation actions, and retarget them to the target humanoid.
Build your robot's motion dataset
If you're training a humanoid to manipulate objects, move, and work in the real world, we capture the human motion that teaches it — and deliver it ready for your pipeline.
Apple Arts Studios — from human motion to humanoid robots.
