01
A friend. A headset.
Your normal walk.
A consenting participant walks while the Quest wearer follows capture guidance. No wearable markers or motion-capture suit.
CAPTURE / META QUEST 3N3 / HUMAN MOVEMENT LAB
See your
stride.
Differently.
You walk. A friend records with a Quest.
Goata gives your movement a new dimension.
BUILT AT HACK THE HILL III RESEARCH PROTOTYPE
MOTION STUDY / 001N3 / CONCEPT
A NEW POINT OF STRIDE
MOVE.
REPLAY.
RETHINK.
REPLAY.
RETHINK.
Illustrative motion. Not a captured person.
01 / THE IDEA
NO MOTION-CAPTURE SUIT.
JUST A DIFFERENT PERSPECTIVE.
Your walk says
a lot about you.
Getting your stride right is hard when you can't see yourself move.
We wanted to change the viewpoint. Put a Quest on a friend, record a walk, and turn that recording into movement you can pause, inspect and walk around.
A mirror gives you one angle.
We wanted the whole picture.
A friend. A headset.
02
Research models.
Working together.
Tracking, pose estimation and body reconstruction turn the recording into fitted motion, with estimated observations linked to source frames.
ANALYSE / LOCAL AI03
Step inside
the replay.
Pause a fitted 3D body, change your viewpoint, scrub a moment or play it backward. Review movement without repeating the performance.
REVIEW / MIXED REALITY02 / THE DEMO FILM
02:17 / PROJECT FILMWatch it work.
The whole picture.
N3 GOATA / PROJECT FILMOpen video in a new tab ↗
03 / INSIDE THE BUILD
Open the repoBuilt on research.
Fuelled by iteration.
Meta Quest. Peer-reviewed reconstruction research. Local GPU inference. A whole lot of algorithmic iteration and GPT tokens.
UNITY / C#META QUEST 3GVHMR / SMPL-XPYTORCH / CUDARTX 3070
FIELD NOTES / LATENCYLess waiting.
Less waiting.
More moving.
We kept the models warm and started processing while the video was still arriving.
MEASURED LOCAL EXPERIMENTBEFORE / WAIT AFTER INPUT119.97s
NOW / WAIT AFTER CAPTURE2.27s*
Read the benchmark conditions
Same 8.008-second public source span. The original run started with the complete input available. The optimized experiment processed 0.5-second chunks during an eight-second paced capture, then needed a median 2.275 seconds to finish and verify the annotated video.
RTX 3070, 640 × 360 analysis, 15 observations per second, FP16 pose/features and warmed resident models. Three post-capture measurements: 2.263, 2.275 and 2.278 seconds. Total time from paced capture start was 10.275 seconds.
This is a local experiment, not a Quest end-to-end measurement. Network delivery, headset loading and watching the video are excluded. It is not a clinical accuracy result or a guarantee for other recordings.
View the measurement summaryPixels are expensive.
Headsets have opinions.
Image compute was slow. Quest integration was janky. We profiled the pipeline, reduced analysis resolution and frame rate, tried precision changes, and stopped reloading the models between jobs.
A clean interface.
In an unfamiliar space.
We built a spatial capture and replay workflow in a medium we hadn't designed for before, with source-linked observations and controls that leave the practitioner in charge.
04 / WHAT'S NEXT
PROD
BABY.
Next comes real-world validation, fresh Quest recordings and the work it takes to earn a place outside a hackathon.