Machine Learning Techniques for Solution of the Inverse Kinematic Problem in Space Manipulators - Master’s Thesis

Aug 2, 2026

Robotics is central to space exploration, enabling reliable execution of tasks difficult for humans under harsh environmental conditions. Achieving efficiency and autonomy in space robotic manipulators requires solving inverse kinematics, especially for redundant configurations with many degrees of freedom, where analytical methods are limited, traditional approaches are inefficient, and singularities occur. The thesis evaluates two machine learning strategies—supervised learning and reinforcement learning—using a feedforward neural network and a Deep Deterministic Policy Gradient (DDPG) agent. Results for 3-DOF validation are extended to 7-DOF and compared, highlighting reinforcement learning as a promising direction for further study.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages66

Reading Time2 h 46 min

AccessFree to download

Introduction

Literature Review

Research Objectives

Thesis Contribution

Robotics Introduction

Direct Kinematics

Inverse Kinematics

Model Description

Machine Learning Overview

Artificial Neural Networks

Deep Learning

Supervised Learning

Reinforcement Learning

Simulation and Training

Supervised Learning Approach

Reinforcement Learning Approach

Analysis and Method Comparison

Machine Learning Techniques for Solution of the Inverse Kinematic Problem in Space Manipulators Tesi di Laurea Magistrale in Space Engineering-Ingegneria Spaziale Author: Luca Marchesotti Student ID: 977060 Advisor: Prof. Mauro Massari Academic Year: 2023-24 i Abstract Robotics plays a crucial role in space exploration as it enables the execution of tasks that are challenging or even impossible for humans to perform in the harsh conditions of space. Ensuring an increasingly safe and reliable future in the space sector requires the achievement of greater efficiency and autonomy in robotic manipulators. One of the primary challenges faced by robotic manipulators, especially those with redundant configurationsand a high number of degrees of freedom, is the resolution of inverse kinematics. This problem arises due to the lack of analytical solutions, the low efficiency of traditional methods, and the presence of singularities. To address this issue, this thesis examines two non-traditional methods that are based on machine learning. Machine learning has experienced exponential growth and application in recent years, making it a promising avenue for solving the problems faced by robotic manipulators. Specifically, the thesis analyzesand compares the supervised learning method and the reinforcement learning method. These methods are applied to a 3 degrees of freedom (DOF) configuration for validation purposes and then extended to a more complex 7 DOF case. The supervised learning method utilizes a feedforward neural network, while the reinforcement learning method employs the Deep Deterministic Policy Gradient Agent. Both methods are evaluated and compared based on their performance in the given configurations. Upon completion of the research, the results obtained from applying these methods in both cases are compared. The analysis highlights reinforcement learning as a potential solution that warrants further investigation. Keywords: Space Manipulators, Inverse Kinematics, Machine Learning, Supervised Learning, Reinforcement Learning, Deep Deterministic Policy Gradient (DDPG) Agent Abstract in lingua italiana La robotica svolge un ruolo fondamentale nell’esplorazione spaziale in quanto consentel’esecuzione di tasks che risultano essere complicate o addirittura impossibili da svolgere per gli esseri umani nelle dure condizioni operative dello spazio. Garantire un futurosempre più sicuro e affidabile nel settore spaziale richiede il raggiungimento di una maggiore efficienza e autonomia dei manipolatori robotici. Una delle sfide principali affrontate per raggiungere questo obbiettivo è la risoluzione della cinematica inversa che, nei manipolatori robotici con configurazioni ridondanti e un elevato numero di gradi di libertà, risulta essere particolarmente complessa. Questo problema sorge a causa della mancanza di soluzioni analitiche, della bassa efficienza dei metodi tradizionali e della presenza di singolarità . Per affrontare questo problema, questa tesi esamina due metodi non tradizionali basati sul Machine Learning (ML) . Il Machine Learning ha registrato una crescita eun’applicazione esponenziale negli ultimi anni, rendendolo uno strumento promettente per risolvere i problemi affrontati dai manipolatori robotici. Nello specifico, la tesi analizzae confronta il metodo di Supervised Learning e il metodo del Reinforcement Learning. Questi metodi vengono applicati inizialmente a una configurazione a 3 gradi di libertà(DOF) per scopi di convalida e quindi estesi a un caso più complesso a 7 DOF. Il metodo di Supervised Learning utilizza una rete neurale feedforward, mentre il metodo di Reinforcement Learning utilizza il Deep Deterministic Policy Gradient Agent. Entrambi imetodi vengono valutati e confrontati in base alle loro prestazioni nelle configurazioni fornite. Al termine della ricerca, vengono confrontati i risultati ottenuti dall’applicazione di questi metodi in entrambe le configurazioni. L’analisi e

FAQs

Why is inverse kinematics difficult for space manipulators with redundant configurations?

Because analytical solutions are largely unavailable, traditional numerical methods are inefficient, and singularities can occur in the solution space.

Which machine learning methods are studied to solve the inverse kinematics problem?

The thesis analyzes and compares supervised learning and reinforcement learning approaches.

How are the supervised and reinforcement learning methods implemented and evaluated?

Supervised learning uses a feedforward neural network, while reinforcement learning uses a Deep Deterministic Policy Gradient (DDPG) agent. Both are tested on a 3-DOF setup for validation and extended to a 7-DOF case, then compared by performance.

Aug 2, 2026

Upload by Seraphina