Tesi etd-03302023-120100
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
PIQUE', FRANCESCO
URN
etd-03302023-120100
Titolo
Model-free control approaches for soft robots
Settore scientifico disciplinare
ING-IND/34
Corso di studi
Istituto di Biorobotica - PHD IN BIOROBOTICA
Commissione
relatore FALOTICO, EGIDIO
Tutor Prof.ssa MENCIASSI, ARIANNA
Tutor Prof.ssa MENCIASSI, ARIANNA
Parole chiave
- soft robotics
- continual learning
- machine learning
- control
- artificial nose
Data inizio appello
09/11/2023;
Disponibilità
parziale
Riassunto analitico
Soft robots promise to revolutionize numerous robotic tasks by virtue of their safety, compliance, dexterity, low cost, and embodied intelligence. However, control of such manipulators remains an open challenge due to the difficulty of developing both accurate and computationally efficient models. Indeed, the materials constituting soft robots exhibit large elastic deformations which endow the robot with virtually infinite degrees of freedom, hysteresis and non linearity which make modeling and thus control difficult.
Avoiding the computation of such models, in the context of control, results in simpler methods which are potentially applicable to different robots without need of any customization.
Model-free control approaches are then defined as those which do not attempt at deriving an analytical or physics-based model of the soft manipulator, but rely either on learning or on a direct relation between sensing and actuation.
Consequently, this thesis presents two model-free control approaches for soft robots. In the first approach, Continual Learning (CL) techniques are applied for the first time in the context of learning-based model-free controllers for soft robots. Indeed, learning-based methods, based on neural networks, can be advantageous in this regard due to the network's ability to capture complex dynamical effects with low computational cost. Continual Learning techniques add further value to these methods by allowing networks to learn from continuously available data without incurring into catastrophic forgetting. Soft robots, which are frequently subject to changes in their dynamic properties, due to material degradation or external interactions, benefit from CL since it allows them to continually adapt to these changes. Notably, in this work this is done without forgetting the control under normal working conditions which can be recovered as soon as the external interactions return to normal. An external weight is used as a proxy of varying dynamics, while elastic weight consolidation (EWC) is used to continuously re-tune a neural network-based controller. It is demonstrated experimentally on a soft robot arm that this method outperforms plain stochastic gradient descent in tracking tasks, in the context of a continuously changing loading condition. It is also shown that the proposed control architecture can improve its performance when exposed to loading conditions already experienced. Secondly, a model-free behavioral approach which relies solely on sensor signals is presented. In particular, an artificial nose based on metal-oxide gas sensors is designed and integrated in a soft robot arm and used for sensing and to drive it towards the source of a volatile organic compound (VOC). Behavioral control does not make use of any system or environment model. In this case the resultant behavior of the soft robot is caused solely by a direct link between sensors and actuators. In particular, the design is such that separate smell concentration readings from different directions around the tip of the robot are possible, and that a one-to-one matching between the sensors’ inputs and the actuators is generated. A simple behavioral control strategy tailored to reach a dynamically varying smell source in the environment is implemented, without any robot modeling. This approach is validated on a two-segment tendon-driven soft robotic arm equipped with the proposed artificial nose. A control strategy for reaching tasks in the case of a stationary smell is also proposed and validated in simulation. Model free approaches can thus represent a valid approach towards controlling soft robots because they can circumvent the complexity of analytic and physics-based modeling of soft manipulators, enabling novel possible tasks in unstructured environments.
Avoiding the computation of such models, in the context of control, results in simpler methods which are potentially applicable to different robots without need of any customization.
Model-free control approaches are then defined as those which do not attempt at deriving an analytical or physics-based model of the soft manipulator, but rely either on learning or on a direct relation between sensing and actuation.
Consequently, this thesis presents two model-free control approaches for soft robots. In the first approach, Continual Learning (CL) techniques are applied for the first time in the context of learning-based model-free controllers for soft robots. Indeed, learning-based methods, based on neural networks, can be advantageous in this regard due to the network's ability to capture complex dynamical effects with low computational cost. Continual Learning techniques add further value to these methods by allowing networks to learn from continuously available data without incurring into catastrophic forgetting. Soft robots, which are frequently subject to changes in their dynamic properties, due to material degradation or external interactions, benefit from CL since it allows them to continually adapt to these changes. Notably, in this work this is done without forgetting the control under normal working conditions which can be recovered as soon as the external interactions return to normal. An external weight is used as a proxy of varying dynamics, while elastic weight consolidation (EWC) is used to continuously re-tune a neural network-based controller. It is demonstrated experimentally on a soft robot arm that this method outperforms plain stochastic gradient descent in tracking tasks, in the context of a continuously changing loading condition. It is also shown that the proposed control architecture can improve its performance when exposed to loading conditions already experienced. Secondly, a model-free behavioral approach which relies solely on sensor signals is presented. In particular, an artificial nose based on metal-oxide gas sensors is designed and integrated in a soft robot arm and used for sensing and to drive it towards the source of a volatile organic compound (VOC). Behavioral control does not make use of any system or environment model. In this case the resultant behavior of the soft robot is caused solely by a direct link between sensors and actuators. In particular, the design is such that separate smell concentration readings from different directions around the tip of the robot are possible, and that a one-to-one matching between the sensors’ inputs and the actuators is generated. A simple behavioral control strategy tailored to reach a dynamically varying smell source in the environment is implemented, without any robot modeling. This approach is validated on a two-segment tendon-driven soft robotic arm equipped with the proposed artificial nose. A control strategy for reaching tasks in the case of a stationary smell is also proposed and validated in simulation. Model free approaches can thus represent a valid approach towards controlling soft robots because they can circumvent the complexity of analytic and physics-based modeling of soft manipulators, enabling novel possible tasks in unstructured environments.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|