Computer Science and Systems Engineering, XXXIII Cycle
PhD defense
Prof. Alberto Bemporad
IMT School for Advanced Studies Lucca, Italy
Dr. Michele Focchi
DLS Lab, Istituto Italiano di Tecnologia, Genova, and
University of Trento, Italy
Assoc. Prof. Mario Zanon
IMT School for Advanced Studies Lucca, Italy
Agile
Strong
source:https://www.youtube.com/embed/tF4DML7FIWk?enablejsapi=1
source:https://youtu.be/pLsNs1ZS_TI
Cassie: https://news.engin.umich.edu/2017/09/latest-two-legged-walking-robot-arrives-at-michigan/
Hexaped: https://www.csiro.au/en/research/
HyQReal: https://dls.iit.it/web/dynamic-legged-systems/robots
sources:
Wheeled robot: https://www.leorover.tech/post/top-3-ways-to-set-a-robot-into-motion
Quadrotor: http://uosdesign.org/designshow2015/ni-myrio-autonomous-quadrotor/
sources:
Robot
Task
Environment
Generate motion plan
Optimal control techniques
Model predictive control with environment adaptation for legged locomotion, IEEE Access, vol. 9, pp. 145710-145727, 2021
Optimization-Based Reference Generator for Nonlinear Model Predictive Control of Legged Robots. Robotics 2023, 12(1), 6, 2023
Model Validation
Model Validation
Trajectory Optimization for Legged Locomotion
Model Validation
Trajectory Optimization for Legged Locomotion
Nonlinear Model Predictive Control
Model Validation
Nonlinear Model Predictive Control
Two-Stage Optimization
Trajectory Optimization for Legged Locomotion
High fidelity model
Rigid Body Dynamics
High fidelity model
Rigid Body Dynamics
Approximate Model
Single Rigid Body Dynamics
Single Rigid Body Dynamics (SRBD)
Discrete-time state space model
Linear time-varying model
Continuous-time state space model
Approximate model
For model validation:
validated by
Validation data: HyQ crawling with the body pitch and yaw
Pitch
yaw
Ground reaction forces for pitch validation
LTV form of the SRBD model
Friction cone and unilateral constraints
Reference \(\mathbf{x}^\mathrm{ref}\) and \(\mathbf{u}^\mathrm{ref}\) generation from heuristic
State \(\mathbf{x}\) and Input \(\mathbf{u}\) tracking term
Uses references (often dynamically infeasible) as linearization trajectory to obtain LTV model
The solution of the LTVOpt may be dynamically infeasible
Initial condition
So the proposed solution is ...
CoM position
CoM velocity
Implementation details:
Ground reaction forces: normal components
Orientation
Angular velocity
[1] J. Koenemann, et al. Whole-body model-predictive control applied to the HRP-2 humanoid. In IEEE/RSJ Int. Conf. Intell. Robot. Syst., pages 3346–3351, 2015.
[2 ] M. Neunert, et al. Whole-Body Nonlinear Model Predictive Control Through Contacts for Quadrupeds. IEEE Robot. Autom. Lett., 3(3):1458–1465,2018.
[3] R. Grandia, et al. Feedback mpc for torque controlled legged robots. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4730–4737, 2019.
[4] A. Herdt, et al. Online walking motion generation with automatic footstep placement. Advanced Robotics, 24(5-6):719–737, 2010.
[5] J. Di Carlo, et al. Dynamic locomotion in the MIT cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1–9,2018.
We use Real-Time Iteration (RTI) scheme for our NMPC.
For online implementation it is desirable:
NMPC runs at \( 25\, \mathrm{Hz} \)
WBC runs at \( 250\, \mathrm{Hz} \)
State estimator runs at \( 500\, \mathrm{Hz} \)
Joint Space PD runs at \( 1\, \mathrm{kHz} \)
Grid map runs at \( 30\, \mathrm{Hz} \)
State \(\mathbf{x}_k \in \mathbb{R}^{12} \)
vector of the robot's CoM position, CoM velocity, base orientation, and angular velocity
Control input \(\mathbf{u}_k \in \mathbb{R}^{12} \)
vector of ground reaction forces
vector of foot locations and contact status
Model parameters \(\mathbf{a}_k \in \mathbb{R}^{16} \)
Decision variables:
Predicted state \(\mathbf{x}^\mathrm{p}=\{\mathbf{x}_{0}, \ldots, \mathbf{x}_N\}\)
Control inputs \(\mathbf{u}^\mathrm{p}=\{\mathbf{u}_{0}, \ldots, \mathbf{u}_{N-1}\}\)
Nonlinear Programming problem
Defined in the CoM frame
Included \(\delta_i\)
Static pallet
Repositioned pallet
Why is it needed?
NMPC
LIP model optimization
QP mapping
+
Remaining state from heuristics
N. Rathod, A. Bratta, M. Focchi, M. Zanon, O. Villarreal, C. Semini, A. Bemporad, “Model Predictive Control with Environment Adaptation for Legged Locomotion”, IEEE Access, vol. 9, pp. 145710-145727, 2021.
A. Bratta, M. Focchi, N. Rathod, and C. Semini, “Optimization-Based Reference Generator for Nonlinear Model Predictive Control of Legged Robots”, Robotics 2023, 12(1), 6, 2023.
Journal papers
A. Bratta, N. Rathod, M. Zanon, O. Villarreal, A. Bemporad, C. Semini, M. Focchi, “Towards a Nonlinear Model Predictive Control for Quadrupedal Locomotion on Rough Terrain”, Italian Institute of Robotics and Intelligent Machines (I-RIM) Conference, 3rd edition, Rome, Italy, 2021.
Conference proceedings and workshops
N. Rathod, M. Focchi, M. Zanon, A. Bemporad, “Optimal control based replanning for Quadruped robots”, Numerical Optimization for Online Multi Contact Planning and Control, Robotics: Science and Systems, Freiburg, Germany, 2019.
N. Rathod, A. Bratta, M. Focchi, M. Zanon, O. Villarreal, C. Semini, A. Bemporad, “Real-time MPC with Mobility enhanced Feature for Legged Locomotion”, Towards Real-World Deployment of Legged Robots, ICRA, Xi’an China, 2021.
[2021 IEEE Access Best Video Award (Part 2)]
[Finalist for the Best Student Paper Award]
Claudio Semini
Head of the Dynamic Legged Systems (DLS) research line
Senior Researcher Tenured - Principal Investigator
Istituto Italiano di Tecnologia, Genova
Angelo Bratta
Postdoctoral Researcher at Dynamic Legged Systems lab, Istituto Italiano di Tecnologia, Genova
Octavio Villarreal
Senior Robotics Research Controls Engineer Dyson
City of Bristol, England, United Kingdom