Why we collect climbing data, how we use it, and the ML pipeline turning movement into injury prevention.
Most climbing injuries come from accumulated stress and poor movement patterns that go undetected until pain shows up. There's almost no large-scale biomechanical data on climbers — sports like running and football have decades of motion capture research, climbing has almost none.
Dynalytix is building the first open dataset of climbing biomechanics. Every video generates 40+ data points per frame — joint angles, speeds, velocities, body positions — all labeled with move types, sensation data, and quality ratings.
This data does two things:
Gives you a detailed breakdown of your own movement patterns — where you're compensating, where you're strong, where you're at risk.
Trains ML models that detect dangerous patterns before they cause injury — not just for you, but for all climbers.
Left/right elbow, shoulder, hip, knee, ankle, plus upper and lower back angles.
Full body tracking with x, y, z coordinates and visibility confidence.
Center of mass, individual landmarks, wrist/hip velocity vectors.
All exported as CSV with frame number and timestamp.
12 supported: Static, Deadpoint, Dyno, Lock-off, Gaston, Undercling, Drop Knee, Heel Hook, Toe Hook, Flag, Mantle, Campus
1–5 form quality score and 0–10 perceived effort per move
9 types (Sharp Pain, Dull Pain, Pop, Unstable, Stretch/Awkward, Strong/Controlled, Weak, Pumped, Fatigue) with body part and intensity
Frame-accurate start/end markers for each move
Video never leaves your browser. MediaPipe JS runs pose extraction entirely client-side — we only store the extracted pose data and your labels, never the video.
Two-stage model architecture: supervised learning on labeled data combined with unsupervised pattern detection.
Supervised Learning
Trains on the collective dataset from all contributors. Using labeled data (move types, quality ratings, sensation tags), it learns:
Transfer Learning + Unsupervised
Once you've uploaded enough sessions, the base model fine-tunes on YOUR data:
├── Supervised: learns from labeled move quality + sensation data
└── Recognizes general "safe" vs "risky" movement patterns
├── Transfer learning: adapts base model to your body
└── Unsupervised: discovers YOUR hidden patterns
├── "Your left shoulder drops 15° more on dynos vs deadpoints"
├── "Your knee valgus increases after minute 40 — fatigue signal"
└── "Your lock-off form degrades when effort > 7"
| Approach | Use Case | Data Needed |
|---|---|---|
| Rule-based engine | Movement assessments with explicit criteria (deep squat) | None — works immediately |
| Supervised ML | Complex movements like climbing | 500–2,000+ labeled examples |
| Unsupervised ML | Hidden pattern detection, anomaly flagging | Accumulates over time per user |
| Hybrid | Rules + supervised + unsupervised | Builds progressively |
We're currently in the data collection phase. The rule-based scoring engine (used in the clinical assessment tool) works today. The ML models are next.
Every contribution makes the system better for everyone.
Contributors upload climbing videos
Pose extraction generates biomechanical data
Contributors label moves + tag sensations
More contributors = more diverse data
Labeled data trains the base model
Better feedback attracts more contributors
Better model = more accurate feedback
Diversity of body types, climbing styles, and skill levels makes the model more robust.
A pattern in your data might prevent someone else's injury.
Syncs to a public GitHub repo (dynalytix-data). Researchers can access it.
Free access to personalized insights when ML models go live.
Help build the future of climbing injury prevention. Every video you label makes the models smarter for everyone.
Start contributing data