Arikex

Motion capture.Now in your pocket.

A metric 3D skeleton from one iPhone. No markers, no suit, no lab.

See how it works ›Sport, medical, games, AI video, previz ›
ANATOMICAL BONE MODEL
16 rigid segments · the app's own rig
BASEBALL · HIT
24 fps · reference motion on the Arikex rig
240 fps capture · 1.2 s LiDAR calibration · 26 keypoints · metric 3D on the phone
Drag to orbit
iPhonecaptures, calibrates and solves the skeleton itself
Apple siliconingests every phone, triangulates, stages the shot
NVIDIA · CUDAtrains the athlete as a 4D splat for bullet time
Why now

Motion capture has been a room.

For thirty years the answer to "how does this body move" was an optical lab: a dozen cameras on a truss, a suit of reflective markers, an afternoon of calibration and a technician to clean the data. It works. It also costs more than a car, never leaves the building, and cannot be pointed at a batter in a cage or a patient in a hallway.

0%

Annual growth of markerless capture through 2030, well ahead of the whole market.

Grand View Research
$25kto$500k

What an optical lab costs, from an entry rig to an in-stadium install.

MoCap Online · KinaTrax
0°

Joint-angle error two iPhones reach against a marker lab. Multi-camera markerless averages 2.3° in gait.

Uhlrich et al. · 2024 meta-analysis
2to8×

Hours of cleanup per hour of optical capture. Calibration alone runs one to two hours.

Studio setup guides, 2025

The optical lab

  • 8 to 40cameras on a truss in a dedicated 3 × 3 × 2.5 m volume
  • 1 to 2 hof wand calibration before the first take, then a marker suit on every subject
  • ~1 mmmarker precision, the bar nobody argues with
  • $500 to $5ka studio day plus a technician to clean the trajectories
  • Next weekthe athlete goes home, the numbers arrive later

Arikex

  • 1 to 4iPhones on tripods or in your hand, wherever the athlete already is
  • 1.2 sLiDAR pre-roll on every take gives the solve true bone lengths, stature and the floor
  • 240 fpscapture with a metric 3D skeleton solved after the take, live preview at 60
  • 0 suitsa 26-keypoint body-and-feet model runs on the phone itself, with a second on-device model for the hands
  • Minutesfrom swing to joint angles, velocities and a coaching read, on the device that shot it
How it works

Four steps. One take.

1 of 4 · Capture

Point the phone. Tap the watch.

Three capture modes on one screen: a live 3D preview, high-speed capture at 30, 60, 120 or 240 fps, and Track, which records the camera's own motion so a handheld shot can be stabilized later. A wrist remote starts and stops the take from the batter's box.

1080p · 240 fpsLiDAR pre-rollWrist remote
2 of 4 · Calibrate

LiDAR measures the body before it moves.

Every take opens with 1.2 seconds of depth. From it Arikex solves bone lengths, stature, distance and a floor plane, then rejects anything that is not a body, like the ghost a mirror returns. High-speed frames inherit the scale: physics does not care about frame rate.

Bone lengths in cmFloor planeAnatomy check
3 of 4 · Solve

2D keypoints become a metric skeleton.

A 26-keypoint model runs on the phone's neural engine, then a whole-take optimizer lifts it into meters: bone-length constancy, a rigid torso, temporal smoothness, ground contact, with LiDAR arbitrating the folds a single camera cannot see. Two phones triangulate instead.

Neural pose, on deviceSequence refiner2-camera triangulation
4 of 4 · Understand

Angles, velocities, and a coach who read them.

Joint angles, hip-shoulder separation, trunk lean and their velocities, timed to the frame. A Deep Analysis pass hands only those numbers to a language model and returns a report where every timestamp scrubs the footage. Video never leaves the phone.

metrics.jsonBVH · USDZAI coaching read
REC 00:01.20
VISION · 240
2D · 26 kpt
SOLO
One capture core, three machines

The phone records. Silicon reasons. CUDA dreams in 4D.

Every take is a folder of plain files: video, depth, keypoints, calibration, camera track. The same analysis core runs on the phone and on current-generation Apple silicon, so a phone can do everything alone or just shoot and send. When a session needs a dynamic Gaussian splat, the workstation hands one command to a CUDA box and brings the frames back.

iPhone

Capture · Solve
  • On-device 2D pose. A 26-keypoint body-and-feet model through the neural engine, with a second model for the hands.
  • Metric 3D on the phone. LiDAR calibration plus a sequence solver. Bone model, 3D viewer, BVH, metrics, AI report.
  • Two phones, one button. A director phone syncs a second over peer-to-peer Wi-Fi and fires both takes together.
  • Wrist remote. Set the tripod, get in the box, tap the watch.

Apple silicon

Ingest · Triangulate · Stage
  • Same engines, bigger machine. Every analysis engine on the phone runs on the workstation through the same code, so a phone can just record and send.
  • Multi-camera truth. Phones find the workstation on the local network and deliver take packages; two views triangulate a real 3D skeleton through a validated biomechanics solver.
  • Stage. Put a take on a floor, add gray props and a camera with a real lens, render a gray-box reference for a video model.

NVIDIA · CUDA

4D splats · Bullet time
  • One command. The workstation converts a session into a per-frame training layout, syncs it to the CUDA box over a private network, trains, and pulls one PLY per frame back.
  • The subject as a 4D splat. Four phones 45° apart give a 135° arc you can freeze, scrub and orbit. Bullet time for a swing, for coaching, never for generation.
  • Nothing else crosses. Every other stage runs on Apple silicon, including a static splat of one frozen instant.
Built to adapt

One skeleton. Five rooms it walks into.

The capture core is the same everywhere: a metric, anatomy-first skeleton with per-segment poses, velocities and confidence. What changes is what you do with it.

A pitching lab in a batting cage.

Hip-shoulder separation, trunk lean, elbow and knee angles and their angular velocities, timed to the frame at 240 fps. A Deep Analysis pass turns the numbers into cues a coach would give, with every timestamp scrubbing the footage. In-stadium systems do this for $500,000 and a camera crew. A cage needs a tripod.

42°Peak separation
240Frames per second
1.08 sTime of peak

Gait analysis in the hallway, not the lab.

A metric skeleton with a real floor gives knee and hip flexion through the gait cycle, stride length in centimeters, cadence and left-right asymmetry, from one phone on a stand. Two phones at front and 45° triangulate for the numbers a clinic can act on. Video stays on the device.

14°Left / right asymmetry
1.24 mStride
108Steps per minute

From take to rig without a suit.

Every solved take exports BVH: a pelvis-rooted hierarchy with calibrated bone offsets, root translation and per-joint rotations, plus USDZ of the posed rig. Drop it on any standard humanoid rig in your engine or DCC. Fingers come from the hand model, feet carry real sole orientation.

25Joints in BVH
60 fpsRotations
0Licensed body models

The reference a video model deserves.

Stage places a take on a LiDAR-scaled floor with gray props and a camera with a real focal length, then renders a gray-box pass at 24 fps with depth, matte and camera JSON. Sent to a current video generation model, the same framing and the same move come back photoreal. Real faces are never uploaded, the mannequin is.

9 sShot at 720p
1:1Framing and move
<3 minReturn

Block the shot where you shot it.

Track mode records the camera's own path with the performance, so a handheld take becomes both an actor and a camera move. Stabilize it to lock the actor to the floor, or import the move as the Stage camera. Export USDZ, BVH and per-frame camera JSON to your DCC, or skip it entirely.

3.84 mCamera travel
115°Sweep
36 mmSensor reference
Accuracy, honestly

We say what a camera can and cannot see.

A single camera cannot tell which way a limb folds in depth. Arikex does not pretend otherwise: LiDAR arbitrates the folds it can measure, the solver enforces what a body can do, and every report carries the take's own self-check, reconstructed stature against the LiDAR measurement. The path to lab-grade numbers is a second phone, and the workstation already walks it.

MethodCamerasJoint-angle error
Optical markers (Vicon, OptiTrack)8 to 40~1 mm position
Markerless multi-camera, gait meta-analysis8+2.3° mean
Theia3D, 16 validation studies8+< 6° RMSE
Two iPhones (OpenCap, marker-validated)24.5° MAE
One iPhone (OpenCap Monocular, 2026)14.8° MAE
Inertial suit (Xsens)06 to 18° RMSE
Arikex, one phone with LiDAR scale1validation shoot planned
Arikex, two phones front + 45°2OpenCap-class pipeline
Sources: Uhlrich et al. 2024 · PMC 2024 meta-analysis · AI in Medicine 2025 · OpenCap Monocular 2026. Arikex figures will be published after a marker-lab shoot.
Arikex mark

Point a phone at it.

Arikex is in private testing with athletes, coaches and clinicians now. If you run a facility, a clinic, a studio or a pipeline that needs motion, we want to point it at yours.

Request a capture session
Arikex · 2026Own rig · own solver · video stays on the deviceiPhone · Apple silicon · NVIDIA CUDA