Motion understanding from video

Record movement.
Understand the body.

One clip is enough — a phone on a tripod, a fixed camera in a treatment room, a handheld follow shot. It comes back as joint angles, reps with pass or fail, stride mechanics, left-right symmetry, range of motion over time, and model-estimated force and stiffness. About five minutes, fully automated. A coach, clinician or teacher sees more of what they were already looking for. Training alone, you see the same reading.

Book a demo → See it in action
6core components
17keypoints per frame
80+metrics, 4 movements
~5 minper clip, automated
ARC ANALYSIS Skeletal tracking · speed & distance overlay

Six components under every measurement

Engine

The engine is six capabilities, each useful on its own. They compose. A sprint, a pull-up, a knee eight weeks after surgery — the same six carry them all. A new movement is a new configuration. The instrument stays the same.

🦴

Monocular pose reconstruction

Seventeen body keypoints per frame from one ordinary camera. No markers, no suit, no depth sensor. Body level: shoulders to ankles, hips to wrists.

Any camera already in the room
🎥

Camera-motion compensation

When the camera pans, pixel displacement is not ground displacement. The engine estimates its own motion between frames and subtracts it, so a handheld follow shot still measures.

Drone, gimbal, phone in hand
🔁

Event detection and counting

A hysteresis state machine marks where each repetition begins and ends. Jitter near the threshold does not double-count. Pass and fail rules stack on top, as many as needed.

Reps, strokes, cycles, gait steps
📐

Scene-geometry self-calibration

Geometry already in the frame carries scale. Track lane lines, court markings, floor tiles. The engine solves for real-world distance from what the scene provides, without tape measures or calibration markers.

Gym floors, clinic tiles, pitch lines
🧬

Joint angle and mechanics

Angles, range of motion, left-right symmetry, leg and vertical stiffness, estimated ground reaction force. Validated biomechanical models turn keypoint trajectories into quantities a coach or clinician already reads.

Rehab progress, load, technique review
⚖️

Judgement as configuration

Measurement is stored in full. The scoring rubric is a configuration file. A federation and a school can read the same clip against different standards, and changing a standard costs no reprocessing.

One archive, many standards
Device makers and application makers license this layer and build on top of it.

Evidence from the field

Case studies

Four movements, every one measured on footage shot in the field — a provincial sports association, a leading national university, a partner testing session, and an evening training run filmed by a running-club captain. Sprint: seventeen athletes, 9.45 m/s peak, ±3% on speed after a single calibration. Pull-ups: 55 repetitions counted, 36 passing strict judgement. Standing long jump: 2.30 to 3.11 m, no floor markers. Distance running: 24 landings measured one by one.

50 m sprint

50 m sprint

The camera follows the runner, so pixel motion contains both the athlete and the operator. The engine separates them, then calibrates against one known result and derives instantaneous speed, stride and contact time.

Components used: Monocular pose reconstruction · Camera-motion compensation · Joint angle and mechanics

17athletes in the sprint cohort
9.45 m/speak instantaneous speed
±3%speed after one calibration
Open the report →
Pull-ups

Pull-ups

Chin-over-bar is easy to see and easy to fake. Kipping and partial range both inflate a count. The state machine segments each repetition, then independent checks decide whether it holds.

Components used: Event detection and counting · Monocular pose reconstruction · Judgement as configuration

55repetitions counted loosely
36passing strict judgement
35%failed on kipping or range
Open the report →
Standing long jump

Standing long jump

Take-off and landing happen in a short stretch of ground with nothing to measure against. Known spacing in the floor markings supplies the scale, and flight time comes from the same keypoint track.

Components used: Scene-geometry self-calibration · Monocular pose reconstruction · Joint angle and mechanics

2.30–3.11 mpilot distance range
0extra calibration markers
20flight and joint metrics
Open the report →
Running form

Distance running · gait and posture

Unlike the timed events, running form has no national standard to score against, so the report gives a portrait rather than a mark: where each foot lands relative to the hip, how much the body rises, trunk and shank angles, and where those values sit among runners. Two-dimensional video repeats well, which makes the same runner's own trend the most reliable comparison.

Components used: monocular pose reconstruction · camera-motion compensation · cross-model validation · foot keypoints

24 stepseach measured, not sampled
184 spmcadence (most runners 155–173)
9.0 cmvertical oscillation (5.5% of height)
See the running-form sample report →

What the system sees

Product

Video in, structured measurement out. A per-athlete report, a scored result against the configured rubric, and the full metric table underneath. We run the engine and hand back the report, the annotated video, and a hosted results page. (shown below with the 50 m sprint.)

Frame-by-frame motion capture

Every joint tracked across the full run, with speed and distance rendered on top frame by frame — the technical detail the eye can't catch, made visible.

Open this athlete's report →

Speed & acceleration profile

A true velocity curve over the whole effort — acceleration, top speed and the point where an athlete starts to fade.

See it in a full report →

Five-dimension performance score

Top speed, speed maintenance, gait symmetry, technique and lower-limb stiffness — scored against biomechanical norms, not opinion.

See it in a full report →

One upload. The rest is automatic.

How it works

Your team does a single thing — record and upload a clip. ARC Motion AI runs the entire pipeline and returns a coach-grade report in minutes. No specialists, no manual analysis.

Your only step
📤

Record & upload

Film an athlete with a phone — even a handheld follow-along shot — and upload the clip. That's the whole ask on your side.

→
⚡ Automated by ARC Motion AI~5 min · zero manual work
01

Capture

Full-body pose reconstructed frame by frame, stabilized against camera motion.

02

Analyze

Speed, cadence, joint angles, symmetry and stiffness — measured and modeled.

03

Report

An individualized, plain-language report with a five-dimension score and training advice.

See a sample report →

Performance-grade analysis from footage anyone can shoot

Our edge

Three problems decide whether single-camera measurement can be trusted: a moving camera, a repetition that only looks complete, and distance without a tape measure. Each was solved for the pilots and each generalises.

🎥

Removing the operator's motion

Even when the camera pans to follow the athlete and the background scrolls, we recover real ground speed and distance — where ordinary methods read close to zero.

🎯

Measured vs. estimated, made honest

Cadence, contact/flight time, joint angles and symmetry are directly measured. Speed and distance are estimated and clearly labeled — and can be calibrated to ±3% with one known result.

⚡

Coach-grade reports at near-zero marginal cost

Raw video to a full individualized report, fully automated — so an assessment that took a specialist an afternoon now scales to an entire cohort.

Start with a phone. Scale when you need to.
No mandatory hardware — no force plates, timing gates or wearables. Accuracy grows with your setup.
1

Phone camera

A handheld clip already reveals technique, cadence and symmetry — a full report in minutes.

Instant
2

+ Stable mount & one known measurement

Fix the camera and give it a real reference — speed and distance calibrate to ±3%.

Calibrated
3

+ Fixed high-frame-rate capture

A dedicated setup gives the full competition-grade biomechanics stack.

Pro

The science inside

Science

Not a black box. Every number is physics — traceable to 100+ peer-reviewed studies, and sharpened by algorithms we engineered ourselves.

01 · Capture

Skeletal reconstruction

Every joint, every frame — 17 body points tracked and stabilized against camera motion, so nothing drifts.

→
02 · Model

Spring–mass physics

Ground force, leg stiffness and stride mechanics from validated running models — physics, not heuristics.

→
03 · Score

Five-dimension verdict

Benchmarked against biomechanical norms — with measured and estimated values always kept distinct.

Camera motion

True speed from a moving camera

A handheld follow shot moves with the athlete. Raw pixel displacement mixes the two. The engine estimates frame-to-frame camera motion from the static background, subtracts it, and leaves the athlete's ground-frame trajectory. Speed then follows from that trajectory scaled by one known result, rather than from raw pixels.

±3% speed after one calibration
Counting integrity

A repetition that holds up

Hysteresis thresholds separate a real repetition from jitter at the turnaround. Three independent checks then run on each one: the chin clears the bar, the arms straighten fully at the bottom, and there is no kipping. A repetition passes when all three agree.

55 loose counts, 36 strict passes
Scene geometry

Distance without a tape measure

Lane lines sit at known spacing. Their projection into the image gives a cross-ratio, which is invariant under perspective. From that invariant the engine recovers metric distance along the ground, so a jump can be measured from a frame with no markers placed.

3.11 m measured, zero markers placed
100+peer-reviewed
studies
20+projects
delivered
80+biomechanical
metrics

Like a mixture of experts, ARC Motion AI combines dozens of specialized, peer-validated methods — each strongest at one thing — and layers algorithms of our own on top.

Spring–mass model · Morin 2005 Sprint Mechanics Score · Bramah 2024 Markerless validation · Van Hooren 2023 Speed determinants · Weyand 2000 …and dozens more

Where the engine applies

Where it applies

The six components do not know which sport they are looking at. A repetition is a repetition, whether it happens at a pull-up bar, on a rehabilitation mat, or in a warehouse aisle. An angle is an angle at the knee, the shoulder, the wrist. The domains below are where those components apply.

🎓

School and university testing

Repetition counts judged against a rubric

  • Counting with hysteresis segmentation
  • Pass and fail rules per movement
  • Rubric configured per institution
🏅

Clubs and academies

Cadence, stride length, contact and flight

  • Per-step timing from one camera
  • Speed requires scene calibration
  • Session-to-session comparison
🏋️

Gyms and fitness centres

Range achieved, tempo, compensation

  • Concentric and eccentric phase timing
  • Left-right symmetry per repetition
  • Depth and lockout flagged per set
🏛️

Federations and regional testing

One rubric applied across every entrant

  • Process data archived beside results
  • Re-score an archive without reprocessing
  • Same standard at every test centre
🩺

Post-operative rehab and physio

Range of motion tracked week over week

  • Left and right compared per session
  • Compensation in hip and trunk
  • Fixed camera in the treatment room
🏠

Home training, remote guidance

Adherence and volume from a phone video

  • No wearables, no equipment
  • Repetition quality alongside volume
  • Weekly log built from uploads
⚽

Team ball sports

Running and jump load across a session

  • Trunk angle on cuts and landings
  • Jump count and flight time
  • Within-set decay in range and speed
🏭

Workplace movement, ergonomics

Repetition count, rhythm, posture exposure

  • Time held in poor trunk posture
  • Lift frequency across a shift
  • Fixed camera on the line

Each domain uses the same six components in a different configuration. Scope stays at body level — trunk, hips, shoulders, knees, ankles, wrists — so what is reported is posture, range, timing and rhythm at joint level.

A large market — validated, and still early

Opportunity

Movement testing is mostly still done by hand. A coach with a stopwatch, a teacher counting a line of students, a physiotherapist with a goniometer and a note. The measurement depends on who is watching, the record is a single number, and the process behind it is discarded. Reading movement from video changes what is kept: the measurement, the frames it came from, and the standard applied to it.

$50BAI in sports, by 203321.6% CAGR
$62BSports technology, by 203021.9% CAGR, from $23B in 2025
$11.6BDigital musculoskeletal care, by 203017.7% CAGR — the rehab lane
$102BGlobal health & fitness clubs→ $235B by 2034, 8.8% CAGR

Market figures: Grand View Research, Fortune Business Insights, Allied Market Research (2024–2025).

27.8%connected-fitness CAGR

At-home training is where the growth is. Virtual and connected fitness is racing from $34B in 2025 toward $312B by 2034 — and every at-home training or rehab device needs a movement-analysis engine behind the camera. That is exactly the layer ARC Motion AI provides.

There is third-party signal for the category. At-home movement rehabilitation companies are publicly listed, private valuations in the sector run into the billions, and a camera-based movement analysis business changed hands for about $285M in 2026. ARC Motion AI sits one level below a finished product: the measurement layer — pose, camera motion, scale, mechanics, rubric — that any movement application needs before it can score anything.

For operators

Deploy it as your platform

Run ARC Motion AI as a branded assessment product for your school system, federation, club network or gym chain — web-based, mobile-ready, and priced to scale to entire cohorts.

For device & app makers

Embed our motion engine

License the biomechanics engine behind a tablet, kiosk or app — the analytics layer for at-home training and rehab hardware, without building a computer-vision team.

Live pilots, real athletes

Proof

Numbers from the two field pilots, covering all three movements. Twenty athletes by code, thirty-one clips, one automated pass per clip from raw video to scored report. Direct measurements and model estimates are reported in separate columns. Open a full sample report →

20athletes across two pilot sites
31clips processed automatically
9.45m/s peak speed, calibrated to one known result
3.11 mlongest jump measured
Sprint — handheld follow shotOpen report →
Pull-ups — one athlete's clip, 14 loose counts, 0 strict passesOpen report →
Long jump — 3.11 mOpen report →
Sprint — lane-calibrated runOpen report →
Pull-ups — one athlete's clip, 14 strict passesOpen report →
Long jump — 3.06 mOpen report →

The team behind it

Team

ARC Motion AI is built by the core team at Arc Collective — based across Silicon Valley and Greater China, with backgrounds spanning AI infrastructure, imaging science, rehabilitation medicine, product and commercialisation.

Team backgrounds: Duke, Stanford, Communication University of China.

The ARC Motion AI vision: a lifelong, continuous record of movement capability for everyone who trains.

David Xue
David Xue
Founder & Chief Executive Officer
Duke. Led AI product lines generating $30M+ annual revenue.
Ti Wu
Ti Wu
Chief Product Officer
Co-founder of Mobike ($2.7B exit).
Lawrence Lou
Lawrence Lou
Chief Investment Officer
100+ Silicon Valley startup investments. Ex-Airbnb / Microsoft.
Shay Luo
Shay Luo
Chief Business Officer
Former Partner at Kearney. Average deal size $30M.
Zefeng Ying
Zefeng Ying
Chief AV & Imaging Scientist
Professor, Communication University of China.
Jinming Yang
Jinming Yang
Chief AI for Pharma & Health
Duke PhD. Cancer therapeutics research.
Jiarong Xing
Jiarong Xing
AI Inference Infrastructure
Founder of KV-Cached. Works with Databricks co-founder Ion Stoica.
Jacky Kwok
Jacky Kwok
Robotics & Embodied AI
PhD at Stanford. Vision-language-action models. Advised by Azalia Mirhoseini and Chelsea Finn.
Get started

One movement is enough to begin

Send a short clip of the movement you care about, with the geometry that is already in the frame. We return the metric table, the segmentation, and a scored result against a rubric you define. Device makers and application makers who want the engine underneath their own product can license this layer, and the integration is done with us.

Book a demo →