Learning Safe Controllers for Motion Generation in Redundant Robots
Apprentissage de contrôleurs surs pour la generation du mouvement des robots redondants
Résumé
One of the key problems in planning and control of redundant robots is the fast gen- eration of controls when multiple tasks and constraints need to be satisfied. In the literature, this problem is classically solved by multi-task prioritized approaches, where the priority of each task is determined by a weight function, or with trajectory optimiza- tion techniques. In this thesis we propose a framework that, through the combination of machine learning and control theory approach, can be efficiently applied both for multi task prioritization and trajectory optimization to automatically find optimal behaviour. First we learn the temporal profiles of the task priorities, represented as parametrized weight functions: we automatically determine their parameters through a constrained stochastic optimization procedure. Then we extend the proposed method for trajectory optimization scenario where we learn the task trajectories for whole-body balancing tasks. In both cases we ensure that the optimized movements are safe and never violate any of the robot and problem constraints. For this purpose we compare three constrained variants of CMA-ES on several benchmarks, among which two are new robotics bench- marks of our design using the KUKA LWR. We retain (1+1)-CMA-ES with covariance constrained adaptation [30] as the best candidate to solve our problems. In order to tackle the limitations of the algorithm described in [30], in this thesis we propose an extension of (1+1)-CMA-ES with Constrained Covariance Adaptation (CCA) that ad- dresses all the issue that affects the learning module of our framework. We show the effectiveness of the proposed framework for the task priority learning on a simulated 7 DOF Kuka LWR and both a simulated and a real Kinova Jaco arm. We compare the performance of our approach to a state-of-the-art method based on soft task prioriti- zation, where the task weights are typically hand-tuned. Then we apply our method on two whole-body experiments with the iCub humanoid robot to show its scalability property. Finally we test our learning framework to the prioritized whole-body torque controller of iCub, to optimize the robot’s trajectory for standing up from a chair.
One of the key problems in planning and control of redundant robots is the fast gen- eration of controls when multiple tasks and constraints need to be satisfied. In the literature, this problem is classically solved by multi-task prioritized approaches, where the priority of each task is determined by a weight function, or with trajectory optimiza- tion techniques. In this thesis we propose a framework that, through the combination of machine learning and control theory approach, can be efficiently applied both for multi task prioritization and trajectory optimization to automatically find optimal behaviour. First we learn the temporal profiles of the task priorities, represented as parametrized weight functions: we automatically determine their parameters through a constrained stochastic optimization procedure. Then we extend the proposed method for trajectory optimization scenario where we learn the task trajectories for whole-body balancing tasks. In both cases we ensure that the optimized movements are safe and never violate any of the robot and problem constraints. For this purpose we compare three constrained variants of CMA-ES on several benchmarks, among which two are new robotics bench- marks of our design using the KUKA LWR. We retain (1+1)-CMA-ES with covariance constrained adaptation [30] as the best candidate to solve our problems. In order to tackle the limitations of the algorithm described in [30], in this thesis we propose an extension of (1+1)-CMA-ES with Constrained Covariance Adaptation (CCA) that ad- dresses all the issue that affects the learning module of our framework. We show the effectiveness of the proposed framework for the task priority learning on a simulated 7 DOF Kuka LWR and both a simulated and a real Kinova Jaco arm. We compare the performance of our approach to a state-of-the-art method based on soft task prioriti- zation, where the task weights are typically hand-tuned. Then we apply our method on two whole-body experiments with the iCub humanoid robot to show its scalability property. Finally we test our learning framework to the prioritized whole-body torque controller of iCub, to optimize the robot’s trajectory for standing up from a chair.
Domaines
Robotique [cs.RO]
Origine : Fichiers produits par l'(les) auteur(s)