Sasan Alibabaei
Contact Body Classification in Human–Robot Collaboration Using Domain Randomization.
Rel. Giuseppe Bruno Averta, Jan Piosik. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Meccanica (Mechanical Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (5MB) | Preview |
Abstract
Use of Transfer Learning inSafe physical interaction is a key requirement in human–robot collaboration. After detecting an external contact, the robot should also determine which part of its structure is involved so that an appropriate reaction can be selected. This thesis investigates the sim-to-real transfer of a seven-class contact-body classifier for a planar 3-RRR parallel robot. All classifier training data are generated in simulation using MuJoCo and MATLAB. A feedforward neural network (FFNN) is trained using twelve physically interpretable features based on the external wrench, actuator torques, and the geometry of the external-force line of action. The main objective is to improve transfer to the physical robot without using physical data for network training or feature alignment.
To this end, physical model parameters are randomized around a reference model, while observation noise and actuator latency are investigated separately
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
