Xiongtao Nie
Implementation and Validation of an Action-Routed Seamless Loop Generation Plug-in for MotionBuilder
Summary
Creating reliable motion loops is a recurring challenge in game animation, virtual production, and motion-capture editing. This project develops and validates Seamless Loop Tool, a Python plug-in for Autodesk MotionBuilder that provides an end-to-end, non-destructive route from motion analysis to FBX export. The tool samples an 18-joint skeleton, extracts 175-dimensional descriptors from overlapping 45-frame windows, and uses a random-forest classifier to route a Take as walk, run, or other. Walk and run motions share a gait-loop detector, while unsupported or uncertain motions remain recoverable through explicit user confirmation.
Candidate loop boundaries are derived from hips-trajectory peaks and bounded autocorrelation, then ranked using endpoint height, Euler rotation, local velocity, duration, and period terms. The processing workflow supports in-place root conversion, full-interval linear endpoint-offset compensation, heading alignment, heuristic foot locking, and sandbox-Take FBX export at 30, 60, 90, or 120 FPS. On 286 clips from a held-out subject, the classifier achieved 96.85% accuracy and a macro-F1 score of 0.9532. Pure NumPy inference reproduced the training-side evaluator with identical labels and a maximum probability error of 4.44 × 10−16, while all 126 automated tests passed. The work substantiates the implementation and code-verification claims, while recognising that external-domain generalisation and measured seam quality, foot-sliding, latency, memory, and artist-productivity gains require further evaluation.
Thesis: Implementation and Validation of an Action-Routed Seamless Loop Generation Plug-in for MotionBuilder