1. Machine learning in legged robotics
- Author
-
Tadić, Tomislav and Ćurković, Petar
- Subjects
Reinforcement Learning Designer ,reinforcement learning ,Deep Network Designer ,učenje s pojačanjem ,Simulink ,SolidWorks ,Environment ,neural networks ,Agent ,TEHNIČKE ZNANOSTI. Strojarstvo ,neuronske mreže ,Okolina ,TECHNICAL SCIENCES. Mechanical Engineering ,Matlab - Abstract
U ovom radu je na početku dan kratak opis strojnog učenja i nekih metoda strojnog učenja kao što su: nadgledano strojno učenje, nenadgledano strojno učenje, polunadgledano strojno učenje i učenje s pojačanjem. Osim toga, ukratko su opisane i umjetne neuronske mreže te genetski algoritmi koji su dijelovi umjetne inteligencije. Nakon toga fokus je stavljen na metodu strojnog učenja pod nazivom “učenje s pojačanjem“. Glavni cilj ovog rada bio je implementirati tu metodu na primjeru modela dvonožnog robota, odnosno pomoću učenja s pojačanjem doći do zakona upravljanja koji će u simulacijskom okruženju pokretati robota, odnosno upravljati njegovim hodom. U radu je opisan cjelokupni postupak od izrade modela robota u SolidWorksu, prebacivanja tog modela u Simulink okruženje (dio Matlaba), prilagodbe tog modela za implementaciju učenja s pojačanjem, treniranja RL (eng. reinforcement learning) Agenta, sve do prikaza hoda modela robota upravljanog istreniranim RL Agentom unutar simulacijskog okruženja kreiranog u Simulink-u. In this paper, a brief description of machine learning and some machine learning methods such as: supervised machine learning, unsupervised machine learning, semi-supervised machine learning and reinforcement learning is given at the beginning. In addition, artificial neural networks and genetic algorithms that are parts of artificial intelligence are briefly described. After that, the focus was placed on the machine learning method called "reinforcement learning". The main goal of this work was to implement that method on the example of a bipedal robot model, that is, by means of reinforcement learning, find the control laws that will move the robot in a simulation environment, that is, manage its gait. The paper describes the entire process of creating a robot model in SolidWorks, transferring that model to the Simulink environment (part of Matlab), adapting that model to implement reinforcement learning, training the RL (reinforcement learning) Agent, up to the presentation of the gait of the robot model controlled by the trained RL Agent within the simulation environment created in Simulink.
- Published
- 2023