arXiv:2503.18674v3 Announce Type: replace 
Abstract: We introduce Human Motion Unlearning and motivate it through the concrete task of preventing violent 3D motion synthesis, an important safety requirement given that popular text-to-motion datasets (HumanML3D and Motion-X) contain from 7\% to 15\% violent sequences spanning both atomic gestures (e.g., a single punch) and highly compositional actions (e.g., loading and swinging a leg to kick). By focusing on violence unlearning, we demonstrate how removing a challenging, multifaceted concept can serve as a proxy for the broader capability of motion "forgetting." To enable systematic evaluation of Human Motion Unlearning, we establish the first motion unlearning benchmark by automatically filtering HumanML3D and Motion-X datasets to create distinct forget sets (violent motions) and retain sets (safe motions). We introduce evaluation metrics tailored to sequential unlearning, measuring both suppression efficacy and the preservation of realism and smooth transitions. We adapt two state-of-the-art, training-free image unlearning methods (UCE and RECE) to leading text-to-motion architectures (MoMask and BAMM), and propose Latent Code Replacement (LCR), a novel, training-free approach that identifies violent codes in a discrete codebook representation and substitutes them with safe alternatives. Our experiments show that unlearning violent motions is indeed feasible and that acting on latent codes strikes the best trade-off between violence suppression and preserving overall motion quality. This work establishes a foundation for advancing safe motion synthesis across diverse applications. Website: https://www.pinlab.org/hmu.

تم تقديم مفهوم نسيان الحركة البشرية لمواجهة تحدي منع توليد الحركات العنيفة في الرسوم المتحركة ثلاثية الأبعاد، وخاصة في مجموعات البيانات مثل HumanML3D وMotion-X، التي تحتوي على نسبة ملحوظة من التسلسلات العنيفة. تهدف هذه المبادرة إلى إنشاء معيار تقييم منهجي لنسيان الحركة من خلال تصفية الحركات العنيفة مع الاحتفاظ بالحركات الآمنة.

Se ha introducido el concepto de Desaprendizaje del Movimiento Humano para abordar el desafío de prevenir la síntesis de movimientos violentos en animaciones 3D, particularmente en conjuntos de datos como HumanML3D y Motion-X, que contienen un porcentaje notable de secuencias violentas. Esta iniciativa tiene como objetivo establecer un marco de evaluación sistemática para el desaprendizaje de movimientos al filtrar los movimientos violentos mientras se retienen los seguros.

Le concept de désapprentissage des mouvements humains a été introduit pour relever le défi de la prévention de la synthèse de mouvements violents dans les animations 3D, en particulier dans des ensembles de données comme HumanML3D et Motion-X, qui contiennent un pourcentage notable de séquences violentes. Cette initiative vise à établir un benchmark d'évaluation systématique pour le désapprentissage des mouvements en filtrant les mouvements violents tout en conservant ceux qui sont sûrs.

The concept of Human Motion Unlearning has been introduced to address the challenge of preventing violent motion synthesis in 3D animations, particularly in datasets like HumanML3D and Motion-X, which contain a notable percentage of violent sequences. This initiative aims to establish a systematic evaluation benchmark for motion unlearning by filtering out violent motions while retaining safe ones.

Human Motion Unlearning

arXiv:2512.11654v1 Announce Type: new 
Abstract: The acquisition cost for large, annotated motion datasets remains a critical bottleneck for skeletal-based Human Activity Recognition (HAR). Although Text-to-Motion (T2M) generative models offer a compelling, scalable source of synthetic data, their training objectives, which emphasize general artistic motion, and dataset structures fundamentally differ from HAR's requirements for kinematically precise, class-discriminative actions. This disparity creates a significant domain gap, making generalist T2M models ill-equipped for generating motions suitable for HAR classifiers. To address this challenge, we propose KineMIC (Kinetic Mining In Context), a transfer learning framework for few-shot action synthesis. KineMIC adapts a T2M diffusion model to an HAR domain by hypothesizing that semantic correspondences in the text encoding space can provide soft supervision for kinematic distillation. We operationalize this via a kinetic mining strategy that leverages CLIP text embeddings to establish correspondences between sparse HAR labels and T2M source data. This process guides fine-tuning, transforming the generalist T2M backbone into a specialized few-shot Action-to-Motion generator. We validate KineMIC using HumanML3D as the source T2M dataset and a subset of NTU RGB+D 120 as the target HAR domain, randomly selecting just 10 samples per action class. Our approach generates significantly more coherent motions, providing a robust data augmentation source that delivers a +23.1% accuracy points improvement. Animated illustrations and supplementary materials are available at (https://lucazzola.github.io/publications/kinemic).

تم تقديم KineMIC (التعدين الحركي في السياق) كإطار للتعلم الانتقالي يهدف إلى تحسين تخليق الإجراءات القليلة العينات لتعرف النشاط البشري (HAR). يتناول هذا الإطار الفجوة الكبيرة بين نماذج النص إلى الحركة العامة ومتطلبات مصنفي HAR الدقيقة، من خلال الاستفادة من التطابقات الدلالية في مساحة ترميز النص من أجل التقطير الحركي.

KineMIC (Kinetic Mining In Context) se ha presentado como un marco de aprendizaje por transferencia destinado a mejorar la síntesis de acciones con pocos ejemplos para el reconocimiento de actividades humanas (HAR). Este marco aborda la significativa brecha de dominio entre los modelos generales de Texto-a-Movimiento (T2M) y los requisitos precisos de los clasificadores HAR, aprovechando las correspondencias semánticas en la codificación de texto para la destilación cinemática.

KineMIC (Kinetic Mining In Context) a été introduit comme un cadre d'apprentissage par transfert visant à améliorer la synthèse d'actions à faible échantillonnage pour la reconnaissance d'activités humaines (HAR). Ce cadre aborde l'écart de domaine significatif entre les modèles Text-to-Motion (T2M) généraux et les exigences précises des classificateurs HAR, en s'appuyant sur des correspondances sémantiques dans l'encodage textuel pour la distillation cinématique.

KineMIC (Kinetic Mining In Context) has been introduced as a transfer learning framework aimed at enhancing few-shot action synthesis for Human Activity Recognition (HAR). This framework addresses the significant domain gap between general Text-to-Motion (T2M) models and the precise requirements of HAR classifiers, leveraging semantic correspondences in text encoding for kinematic distillation.

Human Motion Unlearning

Was this article worth reading? Share it

Humanize AI

GPTHumanizer

sync. labs

Grubby.AI

GPTHuman

AI Humanizer

Ready to build your own newsroom?