This thesis proposes an innovative approach to imitation learning of robotic movements, with the goal of overcoming some typical limitations of standard motion planning algorithms, such as the low flexibility in adapting to different contexts and computational challenges that arise in high-dimensional configuration spaces. The methodology is based on two main pillars: Dynamic Movement Primitives (DMPs), used to encode trajectories into a compact, flexible set of weights that can be easily conditioned on contextual parameters, and Normalizing Flows (NFs), employed as the generative probabilistic model. The framework includes an initial data collection phase through telemanipulation, after which the collected trajectories are encoded into DMPs. A Conditional Normalizing Flow (CNF) is subsequently trained to learn the conditional distribution of the DMP parameters, with the aim of generating new weights consistent with novel contexts. The sampled weights are finally decoded through DMPs into Cartesian trajectories, which are executed by the robot. The approach is initially validated through preliminary experiments designed to test the pipeline, followed by real-world tests that confirm the effectiveness of the method both in simulation and on a UR5e robot. The results demonstrate that CNFs can effectively capture the multimodal nature of trajectories and generate coherent movements in previously unseen contexts. The use of DMPs also ensures the reproduction of smooth and stable trajectories. By implementing different application scenarios, it is shown that the same framework can be adapted with minimal variations to diverse real tasks, highlighting its practicality and versatility. The thesis, therefore, presents and validates an effective and flexible method for planning complex and context-aware robotic trajectories by imitating human movements.
Questa tesi propone un approccio innovativo per l’apprendimento per imitazione di movimenti robotici, con l’obiettivo di superare alcune limitazioni tipiche degli algoritmi standard di motion planning, come la scarsa flessibilità nell’adattarsi a contesti differenti e le difficoltà computazionali che emergono negli spazi di configurazione ad alta dimensionalità. La metodologia si fonda su due pilastri principali: le Dynamic Movement Primitives (DMPs), utilizzate per codificare le traiettorie in un set di pesi compatto, flessibile e facilmente condizionabile a parametri contestuali, e i Normalizing Flows (NFs), impiegati come modello probabilistico generativo. Il framework prevede una prima fase di raccolta dati tramite telemanipolazione, dopo la quale le traiettorie raccolte vengono codificate in DMPs. Viene quindi addestrato un Conditional Normalizing Flow (CNF) per apprendere la distribuzione condizionata dei parametri DMP, con lo scopo di generare pesi coerenti con nuovi contesti. I pesi campionati vengono infine decodificati tramite DMPs in traiettorie Cartesiane, le quali sono eseguite dal robot. L’approccio è validato attraverso esperimenti preliminari, con lo scopo di testare la pipeline, seguiti da test reali che permettono di verificare l’efficacia del metodo sia in simulazione che su un robot UR5e. I risultati dimostrano che i CNFs sono in grado di catturare efficacemente la natura multimodale delle traiettorie e di generare movimenti coerenti con contesti inediti. L'utilizzo delle DMPs garantisce inoltre la riproduzione di traiettorie regolari e stabili. Grazie all’implementazione di diversi scenari applicativi, si dimostra come lo stesso framework possa essere adattato con minime variazioni a diversi task reali, mostrando così la sua praticità d’uso e versatilità. La tesi, quindi, presenta e valida un metodo efficace e flessibile per la pianificazione di traiettorie robotiche complesse e sensibili al contesto che imitano movimenti umani.
Imitation learning of context-aware robot motion primitives with normalizing flows
PERSELLI, CHIARA
2024/2025
Abstract
This thesis proposes an innovative approach to imitation learning of robotic movements, with the goal of overcoming some typical limitations of standard motion planning algorithms, such as the low flexibility in adapting to different contexts and computational challenges that arise in high-dimensional configuration spaces. The methodology is based on two main pillars: Dynamic Movement Primitives (DMPs), used to encode trajectories into a compact, flexible set of weights that can be easily conditioned on contextual parameters, and Normalizing Flows (NFs), employed as the generative probabilistic model. The framework includes an initial data collection phase through telemanipulation, after which the collected trajectories are encoded into DMPs. A Conditional Normalizing Flow (CNF) is subsequently trained to learn the conditional distribution of the DMP parameters, with the aim of generating new weights consistent with novel contexts. The sampled weights are finally decoded through DMPs into Cartesian trajectories, which are executed by the robot. The approach is initially validated through preliminary experiments designed to test the pipeline, followed by real-world tests that confirm the effectiveness of the method both in simulation and on a UR5e robot. The results demonstrate that CNFs can effectively capture the multimodal nature of trajectories and generate coherent movements in previously unseen contexts. The use of DMPs also ensures the reproduction of smooth and stable trajectories. By implementing different application scenarios, it is shown that the same framework can be adapted with minimal variations to diverse real tasks, highlighting its practicality and versatility. The thesis, therefore, presents and validates an effective and flexible method for planning complex and context-aware robotic trajectories by imitating human movements.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_Perselli_Executive_Summary_02.pdf
accessibile in internet per tutti
Descrizione: testo executive summary
Dimensione
9.7 MB
Formato
Adobe PDF
|
9.7 MB | Adobe PDF | Visualizza/Apri |
|
2025_10_Perselli_Thesis_01.pdf
accessibile in internet per tutti
Descrizione: testo tesi
Dimensione
26.35 MB
Formato
Adobe PDF
|
26.35 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243325