Reinforcement learning for complex and dynamic manipulation tasks
Being able to teach complex capabilities, such as opening a dishwasher, to robotic manipulators is a very challenging task, which is often tackled using raw programming. This is an attrac- tive technique for tasks under controlled environments, where variables like relative positions or physical pro...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/428855 |
| Acceso en línea: | https://hdl.handle.net/2117/428855 |
| Access Level: | acceso abierto |
| Palabra clave: | Reinforcement learning Algorithms Manipulators (Mechanism) Aprenentatge per reforç Algorismes Manipuladors (Mecanismes) Àrees temàtiques de la UPC::Informàtica::Robòtica |
| Sumario: | Being able to teach complex capabilities, such as opening a dishwasher, to robotic manipulators is a very challenging task, which is often tackled using raw programming. This is an attrac- tive technique for tasks under controlled environments, where variables like relative positions or physical properties are known and constant between tasks. On the other hand, its perfor- mance degrades in more uncontrolled environments, leading to reduced repeatability and suc- cess rates. Not only that but for some cases not even raw programming is feasible. In this thesis, NVIDIA’s advanced simulation tools, IsaacSim and IsaacLab, are used to create a novel framework that allows the training of reinforcement learning (RL) agents to develop some complex and dynamic manipulation tasks. With the objective of gaining familiarity with the software and iteratively enhance the simula- tion parameters, the framework is first used to explore some basic tasks. Through these pre- liminary tasks, fundamental RL principles were established, gaining insight to create more ad- vanced environments. Not only that but the thesis also explores ways to reduce the simulation to reality (Sim2Real) gap, a critical barrier to deploying simulation-trained models in the real world. Later, some of the studied techniques are used to upgrade the developed framework. The created environment is later used to train a Franka Emika Panda manipulator to open a dish- washer door under a simulated environment. Finally, a connection between simulation and reality is established using ROS2, and put to the test by trying to reproduce one of the simpler trained tasks in a real manipulator. |
|---|