This thesis develops an enhanced version of the kinodynamic RRT* algorithm for motion planning, in which Physics-Informed Neural Control (PINC) networks are integrated into the planning framework. The aim is to generate continuous and near-optimal trajectories that satisfy the system dynamics while maintaining the exploration capabilities of sampling-based methods. The proposed approach combines principles of optimal control with data-driven learning. Two neural models are used: a feedforward neural network predicting the optimal final time and a PINC reconstructing the corresponding state and control trajectory. By embedding the system kinematic model and the optimality conditions derived from Pontryagin’s Maximum Principle into the training process, the networks collectively reproduce the behavior of optimal controllers and act as a learned steering function to efficiently compute feasible and near-optimal local connections between states. This formulation renders numerical optimization or precomputed motion databases unnecessary, drastically reducing memory requirements and computational costs. Each connection can be evaluated quickly, enabling online trajectory generation. PINC-based kinodynamic RRT* is validated through numerical simulations and compared with kinodynamic RRT* MP: the new planner produces smooth, shorter, and dynamically consistent trajectories with respect to the motion-primitives-based variant, while preserving probabilistic completeness and asymptotic optimality. This approach allows high-quality and computationally efficient motion planning suitable for autonomous robotic systems operating in complex environments.
La tesi sviluppa una versione avanzata dell’algoritmo kinodynamic RRT* per la pianificazione del moto, in cui reti neurali fisicamente informate per il controllo (PINC) sono integrate all’interno del pianificatore. L’obiettivo è generare traiettorie continue e quasi ottimali che soddisfino la dinamica del sistema, mantenendo i vantaggi dei metodi basati sul campionamento. L’approccio combina i principi del controllo ottimo con tecniche di apprendimento dai dati. Sono usati due modelli neurali: una rete feedforward, incaricata di prevedere il tempo finale ottimo, e una PINC, che ricostruisce la corrispondente traiettoria di stato e controllo. Integrando nel processo di addestramento il modello cinematico del sistema e le condizioni di ottimalità derivate dal Principio del Massimo di Pontryagin, le reti riproducono il comportamento dei controllori ottimi e consentono connessioni locali tra stati in maniera efficiente. Questa formulazione rende non necessari l’ottimizzazione numerica o i database di traiettorie precomputate, riducendo drasticamente i requisiti di memoria e i costi computazionali. Ogni connessione viene valutata rapidamente, consentendo la generazione online di traiettorie. Kinodynamic RRT* basato su PINC è stato validato mediante simulazioni numeriche e confrontato con kinodynamic RRT* MP: il nuovo pianificatore produce traiettorie più brevi, fluide e dinamicamente consistenti rispetto alla variante basata su primitive di movimento, mantenendo completezza probabilistica e ottimalità asintotica. L’approccio proposto permette una pianificazione del moto di alta qualità ed elevata efficienza computazionale, adatta a robot mobili operanti in ambienti complessi.
PINC-based kinodynamic RRT* for efficient motion planning
Pelagalli, Luca
2024/2025
Abstract
This thesis develops an enhanced version of the kinodynamic RRT* algorithm for motion planning, in which Physics-Informed Neural Control (PINC) networks are integrated into the planning framework. The aim is to generate continuous and near-optimal trajectories that satisfy the system dynamics while maintaining the exploration capabilities of sampling-based methods. The proposed approach combines principles of optimal control with data-driven learning. Two neural models are used: a feedforward neural network predicting the optimal final time and a PINC reconstructing the corresponding state and control trajectory. By embedding the system kinematic model and the optimality conditions derived from Pontryagin’s Maximum Principle into the training process, the networks collectively reproduce the behavior of optimal controllers and act as a learned steering function to efficiently compute feasible and near-optimal local connections between states. This formulation renders numerical optimization or precomputed motion databases unnecessary, drastically reducing memory requirements and computational costs. Each connection can be evaluated quickly, enabling online trajectory generation. PINC-based kinodynamic RRT* is validated through numerical simulations and compared with kinodynamic RRT* MP: the new planner produces smooth, shorter, and dynamically consistent trajectories with respect to the motion-primitives-based variant, while preserving probabilistic completeness and asymptotic optimality. This approach allows high-quality and computationally efficient motion planning suitable for autonomous robotic systems operating in complex environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Pelagalli.pdf
non accessibile
Descrizione: Thesis
Dimensione
7.54 MB
Formato
Adobe PDF
|
7.54 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Pelagalli_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
1.5 MB
Formato
Adobe PDF
|
1.5 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/10589/246311