The aim of this work is to make a robot capable to manage a redundant solution, in order to minimize the cost function associated with a set of tasks. The cost is defined by the user, depending on which aspect is more relevant for the purposes of the final result (e.g. time or precision). Trough reinforcement learning techniques, it will be capable to choose the best solution among the infinite possibilities given by the redundancy of the manipulator. Then the robot will be able to perform different tasks in a three dimensional environment subject to dynamical changes, reacting to different and unpredictable inputs. With the usage of behavior trees and genetic programming techniques, the prioritization of a set of predefined tasks can be learned independently combining them with convex weighted sum and defining the relative parameters. During this learning phase several possible solutions are exploited with operations like crossover and mutation, altering the architecture of the solution. Comparing the result of the cost function, the best resulting algorithm will be chosen. In particular the Baxter robot model was chosen for the problem resolution. The robot is a bimanual one, with 7 degrees of freedom for each arm, and then a mobile base was integrated in order to let it freely move into the environment. The solution of the tasks, like collision avoidance with obstacles, are separately solved and then combined together with null space projector. The simulations were made using Robosuite simulation framework and MuJoCo (MultiJoint dynamics with Contact) engine, through Python programming language.
The aim of this work is to make a robot capable to manage a redundant solution, in order to minimize the cost function associated with a set of tasks. The cost is defined by the user, depending on which aspect is more relevant for the purposes of the final result (e.g. time or precision). Trough reinforcement learning techniques, it will be capable to choose the best solution among the infinite possibilities given by the redundancy of the manipulator. Then the robot will be able to perform different tasks in a three dimensional environment subject to dynamical changes, reacting to different and unpredictable inputs. With the usage of behavior trees and genetic programming techniques, the prioritization of a set of predefined tasks can be learned independently combining them with convex weighted sum and defining the relative parameters. During this learning phase several possible solutions are exploited with operations like crossover and mutation, altering the architecture of the solution. Comparing the result of the cost function, the best resulting algorithm will be chosen. In particular the Baxter robot model was chosen for the problem resolution. The robot is a bimanual one, with 7 degrees of freedom for each arm, and then a mobile base was integrated in order to let it freely move into the environment. The solution of the tasks, like collision avoidance with obstacles, are separately solved and then combined together with null space projector. The simulations were made using Robosuite simulation framework and MuJoCo (MultiJoint dynamics with Contact) engine, through Python programming language.
Learning stack of tasks for robotic mobile manipulation
ADAMI, ALESSANDRO
2023/2024
Abstract
The aim of this work is to make a robot capable to manage a redundant solution, in order to minimize the cost function associated with a set of tasks. The cost is defined by the user, depending on which aspect is more relevant for the purposes of the final result (e.g. time or precision). Trough reinforcement learning techniques, it will be capable to choose the best solution among the infinite possibilities given by the redundancy of the manipulator. Then the robot will be able to perform different tasks in a three dimensional environment subject to dynamical changes, reacting to different and unpredictable inputs. With the usage of behavior trees and genetic programming techniques, the prioritization of a set of predefined tasks can be learned independently combining them with convex weighted sum and defining the relative parameters. During this learning phase several possible solutions are exploited with operations like crossover and mutation, altering the architecture of the solution. Comparing the result of the cost function, the best resulting algorithm will be chosen. In particular the Baxter robot model was chosen for the problem resolution. The robot is a bimanual one, with 7 degrees of freedom for each arm, and then a mobile base was integrated in order to let it freely move into the environment. The solution of the tasks, like collision avoidance with obstacles, are separately solved and then combined together with null space projector. The simulations were made using Robosuite simulation framework and MuJoCo (MultiJoint dynamics with Contact) engine, through Python programming language.File  Dimensione  Formato  

Adami_Alessandro.pdf
accesso aperto
Dimensione
26.05 MB
Formato
Adobe PDF

26.05 MB  Adobe PDF  Visualizza/Apri 
The text of this website © Università degli studi di Padova. Full Text are published under a nonexclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/69281