Climb Analysis

What it measures. How well-structured the centre-of-mass (COM) path is — not how still the climber holds it. Skilled climbers move the COM more, but smoothly and rhythmically; less experienced climbers move it less but jerkily. A higher share of anticipated limb releases is a good sign on vertical terrain; on overhanging terrain anticipatory postural adjustments are naturally reduced or absent, so a lower share there is expected rather than a control deficit.

How it's derived. From the per-frame skeleton (MediaPipe Pose) we track the pelvis/COM and score four things: geometric entropy of the COM path, H = ln(2 × path length / convex-hull perimeter), which is scale-free so a large smooth pendulum scores well and small jitter scores badly — raised entropy can reflect hesitation, correction, exploratory movement or accumulating lower-body fatigue; weight-shift rhythm (regularity of lateral direction reversals); settle quality (how quickly the COM quietens after each move — equilibration); and anticipation (share of limb releases preceded by a COM shift, i.e. anticipatory postural adjustment), read in light of wall inclination. Lateral excursion is reported as a descriptor only and never penalised.

Primary refs: Cordier et al. 1993/1994 (geometric entropy of climbing trajectories); Sibella et al. 2007 (COM trajectories); Quaine & Martin 1999 (equilibrium & anticipatory postural adjustments); Noé 2006 (wall inclination changes anticipatory postural adjustments); Chu 2025 (leg fatigue raises average geometric entropy); Zampagni et al. 2011 (idiosyncratic COM control in experts); Giles et al. 2025 (entropy in the field); MDPI Sensors 23(19):8216 — full entries in About → References

What it measures. Where the hips sit relative to the wall line across the climb, reported as a pattern band: pressed in, workable standoff, or hanging back.

How it's derived. Per-frame horizontal offset between the pelvis and the mean foot position, binned into three bands. "Hips to the wall" is treated as a corrective cue for climbers who hang back, not as something to maximise: elite climbers hold the COM a few centimetres off the wall, which improves the force vector into the feet, keeps holds and footholds in view, and preserves degrees of freedom for the next move. Plastering the wall is flagged as an over-correction rather than rewarded.

Primary refs: Zampagni et al. 2011 (COM distance in expert climbers); Quaine, Martin & Blanchi 1997 (wall reaction forces vs body position); Testa, Martin & Debû 1999 — full entries in About → References

What it measures. Energetic 'cleanliness' — how little wasted motion, regripping, and unnecessary tension shows up in the climb.

How it's derived. Combines pause/rest detection (low-motion segments where the climber could shake out vs hangs unnecessarily), regrip count per hold (how often a hand re-adjusts after first contact), trajectory smoothness of hands and feet (jerk coefficient — third derivative of position, normalised to climb duration), and the model's overall pacing read.

Primary refs: J. Exer. Sci. & Fitness 2019 (jerk coefficient & determinants of climbing performance)

What it measures. Capacity to produce explosive, well-timed force — relevant for dynamic moves, dynos, lock-offs, and steep terrain.

How it's derived. Detects high-acceleration events from the pose stream (peaks in CoM and hand/foot acceleration), counts successful dynamic moves (hand off all holds + reattachment), measures lock-off duration on overhung sections, and asks the multimodal model to score force production qualitatively given the climber's stated grade. Calibrated against the climber's profile (height, weight, ape index).

Primary refs: Frontiers in Sports & Active Living 2025 (anthropometrics); ClimbingCap CVPR 2025 (world-coordinate kinematics)

What it measures. Quality of body positioning and movement choices — drop-knees, flagging, hip-into-the-wall, twist-locks, mantles, etc.

How it's derived. Multimodal pass over the keyframe sequence: the model is prompted to evaluate joint angles, hip orientation relative to the wall, use of opposition and counter-balance, and whether common technique patterns are present where appropriate. Ground-truthed against the on-route hold sequence (only counts technique on holds the climber actually used).

Primary refs: MDPI Sensors 24(19):6479 (activity recognition); Maschek & Schedl 2025 (hold-usage dataset)

What it measures. Precision and intentionality of foot placements — quietness, stability, foot-replacement quality, edging vs smearing accuracy.

How it's derived. Per-foot-contact scoring: foot-position settle time (how long until the foot stops micro-adjusting after first contact), peak vertical-impulse spike at placement (proxy for 'noisy' vs 'quiet' feet), repositioning frequency, and adherence to foot-only holds detected on the route. Each contact is a sample; the score is the trimmed mean across all foot contacts in the climb.

Primary refs: Maschek & Schedl 2025 (contact-based hold detection); EURASIP JASP 2021 (time-series fusion)

What it measures. Whether the climb was delivered at a stable, sustainable rhythm or in bursts separated by stalls — and whether effort was distributed evenly across the route.

How it's derived. Move-to-move interval statistics from the pose track (median inter-move time and its dispersion), combined with a Laban-derived effort read across the climb: the variability of movement speed, weight and flow between successive phases. Low variability alone is not rewarded — the reference is a rhythm appropriate to the terrain.

Primary refs: Draper et al. 2020 (movement/stop segmentation); Laban effort taxonomy as applied to climbing motion — full entries in About → References

What it measures. Whether the current posture already prepares the next hand-move, rather than each move being organised only once the previous one finishes.

How it's derived. A sliding window over the pose sequence tests how well present joint configuration predicts upcoming hand targets, yielding a coupling index and an anticipation depth (how many moves ahead the coupling is detectable). Higher coupling generally indicates more expert, pre-planned movement.

Primary refs: Maselli et al. 2024 (coarticulation in climbing); Seifert et al. (motor-skill coordination) — full entries in About → References

What it measures. How direct versus meandering the hip / centre-of-mass path was through the climb.

How it's derived. Path length compared against the convex-hull perimeter of the same trajectory (H = ln(2 × path length / hull perimeter)), plus a directness ratio of net displacement to travelled distance. Scale-free, so it compares across climbers and framings; interpreted alongside terrain, since traverses and steep sections legitimately raise it.

Primary refs: Cordier et al. 1993/1994; Sibella et al. 2007; Giles et al. 2025 — full entries in About → References

What it measures. Within-path complexity of the pelvis trajectory — amplitude, travelled length, impulse and fractal (Hausdorff) dimension.

How it's derived. Box-counting on the projected pelvis path gives the fractal dimension; amplitude and length are normalised to body scale, and impulse approximates the accumulated pelvic acceleration. Read as a sweet spot rather than a maximum: very low values suggest a stiff, static style, very high values suggest correction and instability. Comparable mainly between attempts on the same route.

Primary refs: Sibella et al. 2007 (pelvis/COM trajectories); Cordier et al. 1994 (fractal characterisation) — full entries in About → References

What it measures. How the climb divides into performatory moves, explorative touches, restorative stops and unproductive stops.

How it's derived. Hand-contact transitions detected from the pose track segment the climb: release-to-next-load-bearing-contact intervals become performatory moves, contacts without loading become explorative, and low-motion phases are split into restorative shake-outs versus stops driven by indecision or error, using motion energy and posture.

Primary refs: Draper et al. 2020 (Front. Psychol., movement/stop taxonomy) — full entries in About → References

What it measures. How long each hand spends loaded versus released, and how balanced that load is left to right.

How it's derived. Per-hand contact intervals summed across the climb give total gripping time, longest continuous grip and the left/right split. Marked left/right imbalance is reported descriptively — it can reflect route layout, a habitual preference, or a functional constraint — and never scored as a fault on its own.

Primary refs: Draper et al. 2020; Fryer et al. (forearm oxygenation & intermittent loading) — full entries in About → References

What it measures. How promptly a climber commits when a viable target and support configuration are available, while separating deliberate information gathering from repeated aborted commitment.

How it's derived. A validation candidate, not a current score: measure time from an eligible target/support state to movement initiation, then timestamp false reaches or hold hovering — repeated hand extension toward a target followed by retraction before a latch. Report raw durations and counts with route, visibility, reachability, fatigue and adaptive context; never label fear, anxiety or confidence from video alone.

Primary refs: Draper et al. 2020 (movement/stop segmentation); Medernach et al. 2021/2025 (decision-making time in bouldering) — full entries in About → References

What it measures. The balance between sampling possible solutions and executing a selected solution during a climb.

How it's derived. A validation candidate, not a current score: count eligible exploratory re-grips, exploratory foot adjustments and non-load-bearing touches, then compare them with first-attempt, load-bearing latches. Present the exploration-to-exploitation ratio beside opportunity counts and route familiarity. Neither a high nor low ratio is inherently better, and camera-derived head orientation remains separate from contact evidence.

Primary refs: Orth et al. 2016/2017 and Seifert et al. 2015 (perception–action exploration in climbing); Draper et al. 2020 (explorative/performatory actions) — full entries in About → References

What it measures. Whether observable movement organization is preserved as cumulative climbing time and likely fatigue increase.

How it's derived. A validation candidate, not a diagnosis of distress tolerance: compare eligible bottom-third and top-third windows for pace, stop structure, contact corrections, path directness and validated smoothness measures. Show each component and the change direction; control for crux location, terrain and route design. A decline may reflect physical fatigue, difficulty or tactics, not reduced composure or cognitive bandwidth by itself.

Primary refs: Walsh et al. 2025 (fatigue, fluidity and hand movements); Draper et al. 2020 (movement/stop structure); Chu 2025 (fatigue and climbing-path entropy) — full entries in About → References

What it measures. Whether the visible head/face orientation remains directed toward the next target immediately before a reach.

How it's derived. A validation candidate, not eye tracking: estimate the pre-reach target-orientation window and dispersion of camera-visible head/face direction. It may be compared with eye-tracked Quiet Eye duration in a validation study, but a standard phone video cannot measure fixation or saccades reliably; reports must say ‘visual-orientation proxy,’ include confidence and allow ‘not assessable.’

Primary refs: Vickers 2007 and Rienhoff et al. 2016 (Quiet Eye); Hacques et al. 2021/2022 (eye-tracked visual control in climbing) — full entries in About → References

What it measures. How distinctly and appropriately movement tempo shifts between easier terrain, micro-rests and crux sequences rather than merely staying constant.

How it's derived. A validation candidate, not a current score: segment eligible climbing phases, derive a robust move-velocity series, discretize or density-model terrain-conditioned velocity states, then calculate entropy and transition structure. Interpret with route sections and rest opportunities; entropy alone is not quality, because both purposeful gear-shifting and disorganized variability can raise it.

Primary refs: Draper et al. 2020 (movement/stop segmentation); Venhorst et al. 2018 (pacing regulation); Cordier et al. 1993/1994 (entropy in climbing) — full entries in About → References

What it measures. Observable self-regulation after a slip, failed sequence or clear execution error: stabilization, reorientation, tactical reset and recommitment.

How it's derived. A validation candidate within mental and tactical performance, not only movement economy: detect a genuine error opportunity, then report time to stable support, time to reorient, whether beta or contact strategy changes, use of an available rest, and outcome of the next attempt. Keep it descriptive until event coding and reliability are established; do not infer emotion or arousal from the behavior alone.

Primary refs: Lazarus & Folkman 1984 and Chesney et al. 2003 (coping and controllability); Cropley et al. 2007 (reflective practice); Draper et al. 2020 (movement/stop segmentation) — full entries in About → References

What it measures. Whether the technique looked correct to an expert eye, scored against a published climbing movement assessment rubric — a different question from the kinematic scores above.

How it's derived. A multimodal pass over the keyframe sequence rates each rubric subscale (e.g. hip/body position, tempo, route reading, commitment) with a short justification. It is a rater view, not a measurement, so it can legitimately diverge from the pose-derived scores.

Primary refs: CM-PAT climbing movement performance assessment rubric — full entries in About → References