social linkedin box blue 32
social facebook box blue 32
social twitter box blue 32
social facebook box blue 32

iit-rbcs-logo-v3

Groups ■ Cognitive Humanoids

Cognitive Humanoids

The Cognitive Humanoids Lab studies how cognitive capabilities, such as the ability of adapting to a continuously evolving environment, can be implement on the iCub humanoid. The interests of the laboratory range from control to human robot interaction passing through mechanical design. In the field of control, the group lead by Francesco Nori studies how the classical control tools (e.g. optimal control and adaptive control) can be ported to novel applications such as involve whole-body postural control. In the field of mechanical design the activities are focused on passive variable impedance actuators with inspiration from human muscles. Finally, in the field of human robot interaction Alessandra Sciutti is studying the idea non-physical interaction and in particular how robots should plan movements so as to convey the same information conveyed in human-human interaction.

Research topics:

ON-LINE WHOLE-BODY DYNAMICS COMPUTATION AND ESTIMATION HYBRID FORCE/POSITION CONTROLLER with passive Variable Stiffness Actuators THE PASSIVE NOISE REJECTION VARIABLE STIFFNESS ACTUATOR (pnrVSA)
alt

On-line Whole-body dynamics computation and estimation

alt

Developing hybrid force/position control capable of adapting to changes in the surrounding environment. 

alt

Novel design and control principles in passive variable impedance actuators: the passive noise rejection variable stiffness actuator (pnrVSA).

LEARNING KINEMATICS BY MEANS OF PRIMITIVES LEARNING HUMAN AND ROBOT ACTION UNDERSTANDING Balancing Strategies in Humanoid Robots with and Stiff and Compliant Actuators
alt

The role of internal models and modularity in the motor adaptation of a reaching task. 

alt

Understanding object properties from the observation of human and humanoid actions to achieve mutual understanding and improve human robot interaction.

alt

Balancing Strategies in Humanoid Robots with and Stiff and Compliant Actuators

EXTERNAL PROJECTS:

CODYCO

Advance the current control and cognitive understanding about robust, goal- directed whole-body motion interaction with multiple contacts

KOROIBOT

Develop humanoids able to walk in a dynamic and versatile fashion in the same way humans do.

SELECTED PUBLICATIONS:


  • Fumagalli M., Ivaldi S., *Randazzo M., *Natale L., *Metta G., *Sandini G. and *Nori F. (2012)
    Force feedback exploiting tactile and proximal force/torque sensing. Theory and implementation on the humanoid robot iCub
    Autonomous Robots, vol. 33, pp. 381-398
  • Metta G., Natale L., Nori F., Sandini G., Vernon D., Fadiga L., Rosander K., Santos-Victor J., Bernardino A. and Montesano L. (2010)
    The iCub Humanoid Robot: An Open-Systems Platform for Research in Cognitive Development
    Neural Networks, vol. 23, (no. 8-9), pp. 1125-1134, 1879-2782
  • *Sciutti A., *Bisio A., *Nori F., *Metta G., *Fadiga L., *Pozzo T. and *Sandini G. (2012)
    Measuring human-robot interaction through motor resonance
    International Journal of Social Robotics, vol. 4, (no. 3), pp. 223-234
alt ON-LINE WHOLE-BODY DYNAMICS COMPUTATION AND ESTIMATION
People involved:

Silvio Traversaro, Andrea Del Prete, Francesco Nori

For efficient control of a humanoid robot complex behavior, is necessary to compute dynamics equations as precisely as possible. However several quantities that are not directly measurable influence the dynamics in a: for  it is possible to integrate the measures from all the different sensors available on the iCub robot (such as 6-axis force/torque sensor, joint level torque sensors, accelerometers, gyroscopes and artificial skin patches).

Multiple sensors of different kind can be exploited, for example:

·         to estimate inertial parameters (inertias, masses, center of mass positions ) taking in account the subset of parameters that is possible to sense from each sensor (embedded 6 axis force/torque sensor or joint torque sensor)

·         to estimate at each instant the external forces, the joint torques, the joint acceleration and velocities fusing together the measures of the available sensors.

ON-LINE WHOLE-BODY DYNAMICS COMPUTATION AND ESTIMATION

Results on arm force dynamics prediction using estimated inertial parameters. 

Fumagalli M., Ivaldi S., *Randazzo M., *Natale L., *Metta G., *Sandini G. and *Nori F. (2012)
Force feedback exploiting tactile and proximal force/torque sensing. Theory and implementation on the humanoid robot iCub
Autonomous Robots, vol. 33, pp. 381-398

*Traversaro S., *Del Prete A., *Muradore R., *Natale L. and *Nori F. (2013)
Inertial Parameter Identification Including Friction and Motor Dynamics
IEEE/RAS International Conference of Humanoids Robotics (HUMANOIDS 2013), Atlanta, GA, USA, October 15-17, 2013

top
alt HYBRID FORCE/POSITION CONTROLLER with passive Variable Stiffness Actuators
People involved:

Francesco Nori, Luca Fiorio, Alberto Parmiggiani, Francesco Romano, Bastien Berret

In the field of robotics there has been a growing interest in bringing robots out of factories in order to operate outside a pre-determined environment, or to work in close contact with humans, e.g. a legged robot walking on uneven terrain, a service robot with elder or disabled people.

It is impossible to accomplish this goal without considering two main issues: safe interaction and adaptation to changes in the environment.

 

This project aims in solving these issues by creating a hybrid force/ position controller capable of adapting to changes in the surrounding environment.

In particular this project will focus on two principal components:

- an adaptive part which continuously estimates a model of the controlled system and of the surrounding environment;

- a motion planner, which takes into account uncertainties in the acquired model.

 

Numerical optimal control, and machine learning tools will be used throughout the project. As a test bed, a newly developed passive variable stiffness actuator (pVSA), presenting high noise rejection capability, will be used.

HYBRID FORCE/POSITION CONTROLLER with passive Variable Stiffness Actuators

Mechanical schematic of the new noise-rejecting passive Variable Stiffness Actuator. Differently from classical agonist-antagonist actuators, in this prototype two additional springs (KPE) are responsible to increase the noise rejection capabilities of this actuator thus allowing an easier control in delayed feedback or openloop control.

*Berret B., *Ivaldi S., *Nori F. and *Sandini G. (2011)
Stochastic optimal control with variable impedance manipulators in presence of uncertainties and delayed feedback
IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2011, pp. 4354-4359, San Francisco 25-30 settembre 2011

*Nori F., *Berret B., *Fiorio L., *Parmiggiani A. and *Sandini G. (2012)
Control of a single Degree of Freedom Noise Rejecting-Variable Impedance Actuator
10th International IFAC Symposiums on Robot Control, pp. 473-478, Dubrovnik, Croatia

top
alt THE PASSIVE NOISE REJECTION VARIABLE STIFFNESS ACTUATOR (pnrVSA)
People involved:

Luca Fiorio, Alberto Parmiggiani, Francesco Romano, Bastien Berret, Francesco Nori

One of the most important features for a system capable of working in uncertain and unstructured environments is reliability. In the last decade roboticisits have started studying mechanical solutions capable of controlling the system structural stiffness by proposing a number of solutions which fall under the broad category of passive Variable Stiffness Actuators (pVSA). In a recent paper (Berret et al., 2011) we observed that none of the available designs seem to reproduce an important characteristic of human muscles, i.e. the ability to open-loop reject disturbances by means of muscle co-activation. Starting from this observation, we propose novel design (Berret et al., 2012) and control (Nori et al., 2012) principles for actuators able to actively regulate the passive noise rejection characteristic (pnrVSA). Through the use of non-linear springs in agonist-antagonist configuration we build a first version (Fiorio et al., 2012) of the actuator, that is currently under test.

The main features of the new actuator:

·         The possibility of finding a close path that connects the frame to the actuator endpoint

·         The capability of co-contracting agonist-antagonist "artificial muscles"

·         The ability of varying the stiffness independently from position 

THE PASSIVE NOISE REJECTION VARIABLE STIFFNESS ACTUATOR (pnrVSA)

Fiorio, L., Parmiggiani, A., Berret, B., Sandini, G., & Nori, F. (2012).

pnrVSA mechanical realization: torque transmission between motors and output joint has been realized employing steel cables Photo by Laura Taverna. 

*Berret B., *Ivaldi S., *Nori F. and *Sandini G. (2011)
Stochastic optimal control with variable impedance manipulators in presence of uncertainties and delayed feedback
IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2011, pp. 4354-4359, San Francisco 25-30 settembre 2011

*Nori F., *Berret B., *Fiorio L., *Parmiggiani A. and *Sandini G. (2012)
Control of a single Degree of Freedom Noise Rejecting-Variable Impedance Actuator
10th International IFAC Symposiums on Robot Control, pp. 473-478, Dubrovnik, Croatia

*Berret B., *Sandini G. and *Nori F. (2012)
Design principles for muscle-like variable impedance actuators with noise rejection property via co-contraction
IEEE-RAS International Conference on Humanoid Robots 2012, Nov.29-Dec.1, 2012. Business Innovation Center Osaka, Japan

*Fiorio L., *Parmiggiani A., *Berret B., *Sandini G. and *Nori F. (2012)
pnrVSA: human-like actuator with non-linear springs in agonist-antagonist configuration
IEEE-RAS International Conference on Humanoid Robots 2012, Nov.29-Dec.1, 2012. Business Innovation Center Osaka, Japan

top
alt LEARNING KINEMATICS BY MEANS OF PRIMITIVES
People involved:

Laura Patanè, Alessandra Sciutti, Bastien Berret, Valentina Squeri, Lorenzo Masia, Giulio Sandini, Francesco Nori

Reaching, apparently one of the simplest human behaviors, is indeed the result of a complex procedure which requires the mapping of sensory signals into the proper muscle activations. This process is thought to rely on internal models, i.e. neural representations of the underlying sensorimotor transformations. In this study we evaluate the role of modularity in the adaptation of these models to different contextual situations. In particular, we tested the prediction that in presence of a modular control, perturbations not compatible with the existing modules should be learned with more difficulty than compatible perturbations. We observed that human subjects not only adapt faster to perturbations compatible with the modules, but they also demonstrate a significant higher capacity of generalization in this condition, thus providing evidence in favor of the adoption of a modular strategy by the central nervous system.

LEARNING KINEMATICS BY MEANS OF PRIMITIVES

1. Experimental setup. On the left, a picture of the experimental setup. On the right, a schema of the two different sets of reaching targets used for the post-exposure phase (yellow dots, in a triangular configuration) and for the training phase (orange dots, in a square configuration).

2. Naïve subject simulation is represented, plotting the error that an untrained subject would do performing each possible transition of the square target-set. Both perturbations tested, one compatible with the modules (intra-modular, solid arrows) and one not compatible (extra-modular, dashed arrows), would theoretically  produce comparable average errors.

3. The average reaching errors for all subjec

*Patanè L., *Sciutti A., *Berret B., *Squeri V., *Masia L., *Sandini G. and *Nori F. (2012)
Modeling kinematic forward model adaptation by modular decomposition
IEEE/RAS-EMBS International Conference on Biomedical Robotics and Biomechatronics, BIOROB 2012, pp. 1252 - 1257, Rome, Italy

Patanè L., Nori F., Berret B., Sciutti A. and Sandini G. (2011)
The role of modularity in learning novel kinematic internal models
21st Annual conference neural control of movement, San Juan, Puerto Rico

top
alt LEARNING HUMAN AND ROBOT ACTION UNDERSTANDING
People involved:

Alessandra Sciutti, Laura Patanè, Francesco Nori, Giulio Sandini

Humans are extremely proficient in their interactions with each other. This efficiency is achieved through an implicit communication exchanged during motion execution. In fact, humans can infer, by observing others’ actions, properties of the manipulated object which are not visible (e.g., its weight). This ability is fundamental during collaboration as it allows the partners to be prepared to act on the object before the hand-over is completed. We have investigated on which action properties such predictive mutual understanding is based, by implementing a simplified lifting action on a humanoid robot and assessing the ability of the human partners to judge object weight from the observation of the robotic lift and to appropriately plan a subsequent action on the same object. Our results show that if robot’s vertical lifting velocity is varied as a function of object weight, subjects reach a performance in weight recognition and action planning comparable to that obtained during human observation, with no extensive learning. These findings suggest that designing automatically understandable robot behaviors could be a possible path to extend human fluid interaction also to human-robot collaborations.

LEARNING HUMAN AND ROBOT ACTION UNDERSTANDING

1. Snapshots of the experimental conditions: an actor (human or robot) lifts bottles of different weights. Subjects are requested to judge the weight and then lift the same object. Photos by Laura Taverna.

2. Velocity modulus  of the demonstrator’s end effector for the various bottles weight in the different experimental conditions. In the robot standard (STD) condition the lifting is planned only by taking into account motion efficacy, in the proportional (PROP) condition instead the understandability of the movement by the human observer is also considered (lifting speed is proportional to object weight, as in human lifting actions).

*Sciutti A., *Patanè L., *Nori F. and *Sandini G. (2013)
Understanding object properties from the observation of the action partner
Human Robot Collaboration Workshop at RSS 2013

*Sciutti A., *Patanè L., *Nori F. and *Sandini G. (2013)
Do humans need learning to read humanoid lifting actions?
Third Joint IEEE International Conference of Development and Learning and on Epigenetic Robotics 2013, Osaka, Japan, August 18-22, 2013

top
alt Balancing Strategies in Humanoid Robots with and Stiff and Compliant Actuators
People involved:

Jorhabib Eljaik, Francesco Nori, Marco Randazzo, Nikolaos Tsagarakis, Zhibin Li

It is a well grounded common believe that the introduction of compliance in robotic systems should provide better contact stability, force control and protection against shock loads (Williamson, 1995). In a joint collaboration with the Department of Advanced Robotics (ADVR) and iCub Facility we are interested in evaluating the benefits of using series elastic actuators (SEA) while performing balancing tasks on humanoid robots and implementing control strategies that allow them to have a safer and more natural interaction with humans and unconstrained environments.

 

Our first efforts have consisted on the use of elastic transmissions in the knees and ankles of the robot. The clever design of these elements however, allows for the possibility of having also traditional rigid joints in the same setup by easily replacing the elastic piece with a rigid one. In this way, fair comparisons of control strategies with and without SEA can be done. To the extent of our knowledge, there are no other robotic platforms with the same advantage. We have been able to see how series elasticity simplifies the role of a balancing controller by low pass filtering the dynamics of the center of pressure (Eljaik et al., 2013), consistently observing more stable balance recovery with SEA in our experimental setups and helping us realize how difficult it is to properly quantify the benefits of their implementation.

 

Our research currently moves towards finding good control strategies for balancing the robot with SEA not only on rigid surfaces but also on soft ones, e.g. a couch, and using multiple contacts to keep whole body balance while performing more complex reaching tasks.

Balancing Strategies in Humanoid Robots with and Stiff and Compliant Actuators

*Eljaik J., *Randazzo M., *Parmiggiani A., *Metta G., *Tsagarakis N. G. and *Nori F. (2013)
Quantitative Evaluation of Standing Stabilization Using Stiff and Compliant Actuators
Robotics, Science and Systems (RSS) 2013, Berlin, Germany, June 24-28, 2013

top