Tesi etd-10022025-085645
Link copiato negli appunti
Tipo di tesi
Corso Ordinario Secondo Livello
Autore
ZILIOTTO, BIANCA
URN
etd-10022025-085645
Titolo
Reinforcement learning-based motion imitation for physiologically plausible musculoskeletal motor control of the arm.
Struttura
Classe Scienze Sperimentali
Corso di studi
INGEGNERIA - INGEGNERIA
Relatori
Tutor Prof. MICERA, SILVESTRO
Relatore Prof. MATHIS, ALEXANDER
Relatore Prof. MATHIS, ALEXANDER
Parole chiave
- motion imitation
- musculoskeletal models
- reinforcement learning
- sensorimotor control
Data inizio appello
11/12/2025;
Disponibilità
parziale
Riassunto analitico
A central objective in sensorimotor control research is to understand how the nervous system generates and coordinates complex movements. Because every internal intention ultimately manifests through muscular activity, movement offers a privileged window into the study of behavior. Numerous computational theories — from optimal feedback control to active inference — have been proposed to explain motor control. However, a key open question is whether these frameworks can scale to the richness and variability of real-world human behavior. One promising direction is to draw from human ethology, using the diversity of natural human movements to develop mechanistic models of sensorimotor control.
During my internship in the A. Mathis Group, I worked toward this goal by extending a computational framework previously developed in the lab: the KINESIS project. KINESIS aims to develop a general sensorimotor control policy through a combination of motion imitation and reinforcement learning. The model is trained using high-quality motion capture data (KIT dataset), which includes locomotor behaviors such as walking, turning, and running. The human body is represented using a realistic musculoskeletal model (Myosuite) with 80 musculotendon units simulated in MuJoCo. Biologically, KINESIS is distinctive in its ability to generate muscle activation patterns that resemble those of real human subjects, outperforming reinforcement-learning–only approaches. This suggests that motion imitation improves task performance while bringing the model’s behavior closer to that of the human neuromuscular system.
My contribution focused on extending this framework to the imitation of upper-limb movements using realistic musculoskeletal models of the arm and hand (Myosuite). This direction enables the exploration of a broader motor repertoire — leveraging the AMASS dataset — and mitigates several challenges associated with lower-limb models, particularly stability issues arising from the absence of abdominal and torso musculature. The next step is to evaluate the proposed imitation-learning model based on the biological plausibility of its generated muscle activation patterns and its learned internal representations.
Finally, I developed a research proposal outlining how this line of work could evolve into my PhD project within the Mathis Group.
During my internship in the A. Mathis Group, I worked toward this goal by extending a computational framework previously developed in the lab: the KINESIS project. KINESIS aims to develop a general sensorimotor control policy through a combination of motion imitation and reinforcement learning. The model is trained using high-quality motion capture data (KIT dataset), which includes locomotor behaviors such as walking, turning, and running. The human body is represented using a realistic musculoskeletal model (Myosuite) with 80 musculotendon units simulated in MuJoCo. Biologically, KINESIS is distinctive in its ability to generate muscle activation patterns that resemble those of real human subjects, outperforming reinforcement-learning–only approaches. This suggests that motion imitation improves task performance while bringing the model’s behavior closer to that of the human neuromuscular system.
My contribution focused on extending this framework to the imitation of upper-limb movements using realistic musculoskeletal models of the arm and hand (Myosuite). This direction enables the exploration of a broader motor repertoire — leveraging the AMASS dataset — and mitigates several challenges associated with lower-limb models, particularly stability issues arising from the absence of abdominal and torso musculature. The next step is to evaluate the proposed imitation-learning model based on the biological plausibility of its generated muscle activation patterns and its learned internal representations.
Finally, I developed a research proposal outlining how this line of work could evolve into my PhD project within the Mathis Group.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|