PeytonT/calisthenics-articulated-pose-v1

Calisthenics articulated pose v1

Browser observation-to-body model for CaliCombos. About 17.5 MB of weights/topology. Nine temporal observations produce a shared 325-value MHR state. The portable decoder returns 127 skeleton joints, 70 landmarks and an 8,471-vertex humanoid. Body, hands, spine, proportions and camera travel together. This is MHR, not an SMPL-X checkpoint.

The input is body/hand pose observations with a camera scaffold. This model does not replace the image detector, establish physical contact, or measure metric depth. Missing detections remain gaps. Runtime sharing of shape/focal across continuity segments is included in training losses. The output coordinates are camera x-right, y-down, z-away.

Refinement includes source-reviewed separated legs in the final step, straight-arm priors, tracked hand-grip proxies and per-joint motion-aware fitting. Contact labels are image proxies. The portable decoder omits the teacher surface-corrective network.

Evidence and limits

On the same development/training clips: Golden Temple passes all 39 reviewed regions with usable browser inputs (39/64 total); the other 25 lack usable input. Fingertip passes 16/16. Previously passing regions are preserved. These are reconstruction/transfer checks, not held-out accuracy. No physical-phone test, calibrated anatomical 3D validation, automatic ground-contact validation or broad skill generalization has been established. The model must be presented as an estimate, with editable annotations and no inferred joints relabeled as measured observations.

The ONNX student and body decoder execute in browser WASM and are verified against native output. The application downloads commit-pinned, SHA-256-verified assets and caches locally. Clip pixels stay on the device. No videos, per-frame pose data or credentials are included here.

Use both manifests together; the pose manifest pins the exact decoder manifest hash. See validation.json and upstream LICENSE-MHR.txt / LICENSE-SAM.txt. MHR geometry and SAM 3D Body fits are acknowledged.