We Just Released 10 Hours of Free Human Motion Capture Data for Robotics — Here's What's Inside

Training a humanoid robot to move like a human needs one thing before anything else: real human motion, captured accurately and labelled cleanly. That data is hard to find, and most of what's freely available is small, noisy, or stripped of the detail that actually matters.

So we did something about it. Apple Arts Studios has released a free 10-hour sample of our human motion capture dataset on Hugging Face — real, production-grade optical mocap you can download and test in your own pipeline today, no strings attached.

You can find it here: huggingface.co/datasets/Appleartsstudios/Motion_Dataset

What's actually in it

This isn't scraped video or rough IMU data. Every motion in this dataset was captured on our optical stage using an OptiTrack PrimeX 41 system at 120 Hz, on a 53-point skeleton that includes full finger articulation — the hand detail most motion datasets leave out.

The sample contains 2,584 files across roughly 10 hours — 5 hours of original captured motion plus 5 hours of mirror-augmented versions. Every single file ships with complete metadata: the action name and category, performer height, weight, age and gender, capture frame rate, skeleton definition, clip duration and frame count, and whether the file is an original or a mirror. Nothing is unlabelled, and nothing is guesswork.

Two formats, ready for your pipeline

Every motion comes in two forms so you can drop it straight into whatever you're building:

SOMA Uniform BVH at 30 FPS — clean, retarget-friendly skeletal animation that works across standard motion pipelines.

Unitree G1 CSV — motion already mapped to the Unitree G1 humanoid, so robotics teams working with that platform can start immediately.

Each clip is trimmed to the action itself, separated per performer and per variation, with T-pose and idle removed. It's built to be trained on, not cleaned up first.

What it covers

The sample spans locomotion and manipulation categories — the movement foundations robots need to learn: walking and turning, crouching, balance, stairs and ramps, object interaction, and manipulation. It's a slice of a much larger library covering the full range of everyday human movement.

Why we're giving it away

Simple: we want robotics and AI teams to try the real thing before they commit. Talking about data quality is cheap. Letting you download it, load it, and see how it behaves in your own model is honest. If it works for you, the full library — over 130 hours of cleaned, training-ready clips plus hundreds of hours of full-length raw capture sequences — is available under commercial licence, and we also capture custom motion to order.

But the sample is free, and it's genuinely useful on its own. Download it, train on it, and tell us what you think.

Who we are

Apple Arts Studios is India's largest motion capture studio, with over fifteen years of experience delivering performance capture for feature films, AAA games, and VFX productions worldwide. We build human motion datasets for the companies teaching machines to move — humanoid robotics, imitation learning, and embodied AI.

Try the free dataset: huggingface.co/datasets/Appleartsstudios/Motion_Dataset License the full library or commission custom capture: hello@appleartsstudios.com · www.appleartsstudios.com