Learning and control with movement primitives in multiple - PowerPoint PPT Presentation

Learning and control with movement primitives in multiple coordinate systems Sylvain Calinon Robot Learning & Interaction Group Idiap Research Institute Martigny, Switzerland

RIBA healthcare robot, RIKEN i-limb prosthetics, TouchBionics Humanoid diver, Stanford Robonaut, NASA

Learning from demonstration as an intuitive interface to transfer skills to robots

+ Statistical learning dynamical systems We look for a modular representation of movements and skills that can learn from wide-ranging data , that can adapt to new situations in a fast manner, that can exploit the robot embodiment , and that is robust to perturbations .

Model predictive control (MPC) Track path! Do it smoothly! System plant How to solve this cost function? • Pontryagin’s maximum principle state variable (position+velocity)  Riccati equation control command (acceleration) • Dynamic programming tracking weight matrix • Linear algebra with stacked control weight matrix vectors [Tanwani and Calinon, IEEE RA-L 1(1), 2016]

Model predictive control (MPC) Track path! Do it smoothly! System plant [Tanwani and Calinon, IEEE RA-L 1(1), 2016]

Model predictive control (MPC) Sharing of synergies with: [Tanwani and Calinon, IEEE RA-L 1(1), 2016]

Model predictive control (MPC) Transition and state duration  Minimal intervention  Safe/compliant robot controller Stepwise sequence with: [Tanwani and Calinon, IEEE RA-L 1(1), 2016]

Transfer of controllers from demonstration Demonstration Reproduction Holding a cup horizontally coordination Bimanual [Calinon, Intelligent Service Robotics 9(1), 2016]

Extension to multiple coordinate systems  in a new situation… Coordinate system 1: This is where I expect data to be located! Coordinate system 2: This is where I expect data to be located!  Product of linearly transformed Gaussians [Calinon, Bruno and Caldwell, ICRA’2014][Calinon, HFR’2016]

MPC considering multiple coordinate systems Track path in coordinate system j Do it smoothly! New position and 2 2 orientation of coordinate 2 2 systems 1 and 2 2 Two candidate 1 1 coordinate systems (P=2) Set of demonstrations Reproduction in new situation [Calinon, Bruno and Caldwell, ICRA’2014][Calinon, HFR’2016]

MPC considering multiple coordinate systems Track path in coordinate system j Do it smoothly! In many robotics problems, the parameters describing the task or situation can be recast as some form of coordinate systems or locally linear transformations 2 1 [Calinon, Bruno and Caldwell, ICRA’2014][Calinon, HFR’2016]

MPC considering multiple coordinate systems Track path in coordinate system j Do it smoothly! In many robotics problems, the parameters describing the task or situation can be recast as some form of coordinate systems or locally linear transformations  Learning of a controller that adapts to new situations while regulating its gains according to the precision and coordination required by the task [Calinon, Bruno and Caldwell, ICRA’2014][Calinon, HFR’2016]

Adaptation of movements to different shapes Candidate coordinate system [Calinon, Alizadeh and Caldwell, IROS’2013]

Bimanual coordination and co-manipulation [Rozo et al., IROS’2015] [Silverio et al., IROS’2015] Dr Leonel Rozo João Silvério [Rozo et al., IEEE T-RO 32(3), 2016]

Joint space constraints & tasks prioritization Task parameters as Jacobian operators Demonstration Reproduction Joint space Task space on COMAN on WALKMAN (configuration space) (operational space) Left hand priority Task parameters as null space projection Right hand priority operators [Silverio, Calinon, Rozo and Caldwell (submitted)] [Calinon, ISRR’2015]

Learning from demonstration can be applied to various forms of robots and applications (2012-2015) (2015-2018) (2015-2018)

Continuous MPC for continuum robots arm index s s=1 s=0 Stiff Compliant [Bruno, Calinon, Malekzadeh and Caldwell, ICIRA, LNCS 9246, 2015]

Flexible robots in minimal invasive surgery Insertion Retraction Collaboration letting the surgeon control the tip, while the robot exploits the remaining degrees of freedom that do not interfere with the control of the tip [Bruno, Calinon and Caldwell, Autonomous Robots, 2016]

Robotic dressing assistance Dressing assistance for: - Putting on a coat - Putting on shoes  Requires to extend movement primitives to reaction, force and impedance primitives

Telemanipulation with underwater robot Same model (TP-HSMM) used for classification on the one side, and synthesis on the other side, with local adaptation to the task parameters Onshore Offshore

Telemanipulation with underwater robot Does not need correction Needs correction Being skillful = exploiting variability and correlation Recognition & synthesis of motion primitives  Semi-autonomous teleoperation as a form of human-robot collaboration [Tanwani and Calinon, IEEE RA-L 1(1), 2016]

Conclusion • Model predictive control (MPC) can be smoothly combined with learning from demonstration (LfD) for both planning and control problems • The parameters of the cost function in MPC can be learned from demonstration, with weights as full precision matrices , instead of predefining those manually as diagonal matrices • LfD can be extended to the consideration of multiple coordinate systems and to the learning of controllers  Minimal intervention strategy that is safer for the users • Extending LfD to collaborative skills and teleoperation provides new perspectives in shared control by exploiting the learned task variations and task synergies to cope with perturbations

Source codes: http://www.idiap.ch/software/pbdlib/ Contact: sylvain.calinon@idiap.ch http://calinon.ch Collaborators at Idiap: Dr Ioannis Havoutis Ajay Tanwani Emmanuel Pignat Noémie Jaquier Photo: Basilio Noris