[PPT] - Control of redundant robots using learned models: an operational PowerPoint Presentation

SLIDE 1

Control of redundant robots using learned models: an operational space control approach

Presented at the IEEE/RSJ International Conference on Intelligent Robots and Systems October 12, 2009 / St Louis - USA

by Camille Salaün, Vincent Padois (vincent.padois@upmc.fr), Olivier Sigaud

Université Pierre et Marie Curie Institut des Systèmes Intelligents et de Robotique (CNRS UMR 7222)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 1 / 14

SLIDE 2

Introduction

◮ Robotics applications: moving from the industrial domain to the service domain (e.g. [1]) Induces increasing complexity

Robots : complex structures, more actuators, more sensors Unstructured / Unknown environments

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 2 / 14

SLIDE 3

Introduction

◮ Robotics applications: moving from the industrial domain to the service domain (e.g. [1]) Induces increasing complexity

Robots : complex structures, more actuators, more sensors Unstructured / Unknown environments

Robustness and adaptivity are required: learning is part of the solution

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 2 / 14

SLIDE 4

Introduction

◮ Robotics applications: moving from the industrial domain to the service domain (e.g. [1]) Induces increasing complexity

Robots : complex structures, more actuators, more sensors Unstructured / Unknown environments

Robustness and adaptivity are required: learning is part of the solution That does not mean throwing away model-based robotics control →it can provide a sound framework and a good starting point

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 2 / 14

SLIDE 5

Introduction

◮ Robotics applications: moving from the industrial domain to the service domain (e.g. [1]) Induces increasing complexity

Robots : complex structures, more actuators, more sensors Unstructured / Unknown environments

Robustness and adaptivity are required: learning is part of the solution That does not mean throwing away model-based robotics control →it can provide a sound framework and a good starting point General goal of this work Propose a methodology to combine learning methods and model-based control ֒ → The work in this paper illustrates a methodology to do so

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 2 / 14

SLIDE 6

Introduction

Outline of this presentation

1

Introduction

2

Background in Operational Space Control

3

Learning strategy and learning tools

4

Simulated experiments description

5

Results and Analysis

6

Conclusions and Perspectives

7

References

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 3 / 14

SLIDE 7

Background in Operational Space Control

◮ Model-based control techniques are mostly based on the knowledge of the model(s) relating the joint space (dimension n) to the task space (dimension m) [2].

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 4 / 14

SLIDE 8

Background in Operational Space Control

◮ Model-based control techniques are mostly based on the knowledge of the model(s) relating the joint space (dimension n) to the task space (dimension m) [2]. These mapping can be described at two different levels Geometric: ξ = h (q) → non-linear, hard to inverse Velocity kinematics: ˙ ξ = J (q) ˙ q

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 4 / 14

SLIDE 9

Background in Operational Space Control

◮ Model-based control techniques are mostly based on the knowledge of the model(s) relating the joint space (dimension n) to the task space (dimension m) [2]. These mapping can be described at two different levels Geometric: ξ = h (q) → non-linear, hard to inverse Velocity kinematics: ˙ ξ = J (q) ˙ q The relation between forces and accelerations is also important Dynamics in the joint space: Γ = A (q) ¨ q + b (q, ˙ q) + g (q) + ǫ (q, ˙ q) − Γext Dynamics in the task space: f = Λ (q) ¨ ξ + µ (q, ˙ q) + p (q) + ǫ′ (q, ˙ q) − f ext [3]

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 4 / 14

SLIDE 10

Background in Operational Space Control

◮ Model-based control techniques are mostly based on the knowledge of the model(s) relating the joint space (dimension n) to the task space (dimension m) [2]. These mapping can be described at two different levels Geometric: ξ = h (q) → non-linear, hard to inverse Velocity kinematics: ˙ ξ = J (q) ˙ q The relation between forces and accelerations is also important Dynamics in the joint space: Γ = A (q) ¨ q + b (q, ˙ q) + g (q) + ǫ (q, ˙ q) − Γext Dynamics in the task space: f = Λ (q) ¨ ξ + µ (q, ˙ q) + p (q) + ǫ′ (q, ˙ q) − f ext [3] Common approaches in model-based control often rely on:

An error signal in the task space The inversion of the joint space to task mapping The use of the dynamics equation to compute the forces to apply to the robot

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 4 / 14

SLIDE 11

Background in Operational Space Control

◮ Model-based control techniques are mostly based on the knowledge of the model(s) relating the joint space (dimension n) to the task space (dimension m) [2]. These mapping can be described at two different levels Geometric: ξ = h (q) → non-linear, hard to inverse Velocity kinematics: ˙ ξ = J (q) ˙ q The relation between forces and accelerations is also important Dynamics in the joint space: Γ = A (q) ¨ q + b (q, ˙ q) + g (q) + ǫ (q, ˙ q) − Γext Dynamics in the task space: f = Λ (q) ¨ ξ + µ (q, ˙ q) + p (q) + ǫ′ (q, ˙ q) − f ext [3] Common approaches in model-based control often rely on:

An error signal in the task space The inversion of the joint space to task mapping The use of the dynamics equation to compute the forces to apply to the robot

◮ In this work we focus on the velocity kinematics level (dynamics is assumed to be known)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 4 / 14

SLIDE 12

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 13

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 14

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆

3 cases have to be distinguished

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 15

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆

3 cases have to be distinguished

Fully constrained (n = m): one exact solution, J (q)♯ = J (q)−1

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 16

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆

3 cases have to be distinguished

Fully constrained (n = m): one exact solution, J (q)♯ = J (q)−1 Over constrained (n < m): no exact solution, minimum error if J (q)♯ is a (weighted) pseudo-inverse [4, 5]

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 17

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆ + NJ ˙

q0 3 cases have to be distinguished

Fully constrained (n = m): one exact solution, J (q)♯ = J (q)−1 Over constrained (n < m): no exact solution, minimum error if J (q)♯ is a (weighted) pseudo-inverse [4, 5] Redundant (n > m): ∞ of sol., minimum norms if J (q)♯ is a (weighted) pseudo-inverse

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 18

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆ + NJ ˙

q0 3 cases have to be distinguished

Fully constrained (n = m): one exact solution, J (q)♯ = J (q)−1 Over constrained (n < m): no exact solution, minimum error if J (q)♯ is a (weighted) pseudo-inverse [4, 5] Redundant (n > m): ∞ of sol., minimum norms if J (q)♯ is a (weighted) pseudo-inverse

Redundancy can be used to locally optimize a cost function or to hierarchically realize a secondary task [6, 7, 8]: ˙ q = J♯

1 ˙

ξ

⋆ 1 + (J2NJ1)♯ ˙

ξ

⋆ 2 − J2J♯ 1 ˙

ξ

⋆ 1

.
V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 19

Background in Operational Space Control

Inverse Velocity Kinematics

◮ Problem : given a desired task space velocity ˙ ξ

⋆, compute the corresponding ˙

q Assuming singularity free configurations, the solution can be written as ˙ q = J (q)♯ ˙ ξ

⋆ + NJ ˙

q0 3 cases have to be distinguished

Fully constrained (n = m): one exact solution, J (q)♯ = J (q)−1 Over constrained (n < m): no exact solution, minimum error if J (q)♯ is a (weighted) pseudo-inverse [4, 5] Redundant (n > m): ∞ of sol., minimum norms if J (q)♯ is a (weighted) pseudo-inverse

Redundancy can be used to locally optimize a cost function or to hierarchically realize a secondary task [6, 7, 8]: ˙ q = J♯

1 ˙

ξ

⋆ 1 + (J2NJ1)♯ ˙

ξ

⋆ 2 − J2J♯ 1 ˙

ξ

⋆ 1

.

Here, 2 cases have to be distinguished

Compatible tasks (n ≥ m1 + m2): the two tasks can be perfectly achieved Incompatible (n < m1 + m2) : task 2 cannot be perfectly achieved but has minimum error

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 5 / 14

SLIDE 20

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12]

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 21

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12] Our claim At the velocity level, one should rather learn forward mappings ([13]) for each task because ...

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 22

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12] Our claim At the velocity level, one should rather learn forward mappings ([13]) for each task because ...

... directly learning inverse mappings leads to a loss of information about the redundant nature of the system

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 23

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12] Our claim At the velocity level, one should rather learn forward mappings ([13]) for each task because ...

... directly learning inverse mappings leads to a loss of information about the redundant nature of the system ... a priori task combination requires to (re)learn everything again when modifying the task combination

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 24

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12] Our claim At the velocity level, one should rather learn forward mappings ([13]) for each task because ...

... directly learning inverse mappings leads to a loss of information about the redundant nature of the system ... a priori task combination requires to (re)learn everything again when modifying the task combination

Possible drawbacks of this approach

the matrix inversion process may lead to the amplification of learning error the learning errors may lead to a loose tasks hierarchy and decoupling

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 25

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12] Our claim At the velocity level, one should rather learn forward mappings ([13]) for each task because ...

... directly learning inverse mappings leads to a loss of information about the redundant nature of the system ... a priori task combination requires to (re)learn everything again when modifying the task combination

Possible drawbacks of this approach

the matrix inversion process may lead to the amplification of learning error the learning errors may lead to a loose tasks hierarchy and decoupling

We incrementally learn the different tasks Jacobians using sets of inputs and outputs (q, ξ)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 26

Learning strategy and learning tools

Learning strategy and tools

In the literature, some of the proposed approaches directly try to learn inverse mappings [9, 10] and also statically try to solve the redundancy using an extended Jacobian ([11]) to combine tasks [12] Our claim At the velocity level, one should rather learn forward mappings ([13]) for each task because ...

... directly learning inverse mappings leads to a loss of information about the redundant nature of the system ... a priori task combination requires to (re)learn everything again when modifying the task combination

Possible drawbacks of this approach

the matrix inversion process may lead to the amplification of learning error the learning errors may lead to a loose tasks hierarchy and decoupling

We incrementally learn the different tasks Jacobians using sets of inputs and outputs (q, ξ) To do so we use the Locally Weighted Projection Regression algorithm (LWPR) [14]

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 6 / 14

SLIDE 27

Learning strategy and learning tools

Learning strategy and tools

LWPR is an incremental function approximator which provides accurate approximation in very large spaces in O(k)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 7 / 14

SLIDE 28

Learning strategy and learning tools

Learning strategy and tools

LWPR is an incremental function approximator which provides accurate approximation in very large spaces in O(k) Learning of a task Jacobian Ji: modeli = LWPRlearn (q, ξi) Prediction of the task Jacobian Ji: ˆ ξ1, ˆ Ji

= LWPRpredict (q, modeli)
V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 7 / 14

SLIDE 29

Learning strategy and learning tools

Learning strategy and tools

LWPR is an incremental function approximator which provides accurate approximation in very large spaces in O(k) Learning of a task Jacobian Ji: modeli = LWPRlearn (q, ξi) Prediction of the task Jacobian Ji: ˆ ξ1, ˆ Ji

= LWPRpredict (q, modeli)

Control using the learned model: ˙ q = ˆ J♯

1 ˙

ξ

⋆ 1 + ˆ

J2Pˆ

J1

♯ ˙

ξ

⋆ 2 − ˆ

J2ˆ J♯

1 ˙

ξ

⋆ 1

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 7 / 14

SLIDE 30

Learning strategy and learning tools

Learning strategy and tools

LWPR is an incremental function approximator which provides accurate approximation in very large spaces in O(k) Learning of a task Jacobian Ji: modeli = LWPRlearn (q, ξi) Prediction of the task Jacobian Ji: ˆ ξ1, ˆ Ji

= LWPRpredict (q, modeli)

Control using the learned model: ˙ q = ˆ J♯

1 ˙

ξ

⋆ 1 + ˆ

J2Pˆ

J1

♯ ˙

ξ

⋆ 2 − ˆ

J2ˆ J♯

1 ˙

ξ

⋆ 1

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 7 / 14

SLIDE 31

Learning strategy and learning tools

Learning strategy and tools

LWPR is an incremental function approximator which provides accurate approximation in very large spaces in O(k) Learning of a task Jacobian Ji: modeli = LWPRlearn (q, ξi) Prediction of the task Jacobian Ji: ˆ ξ1, ˆ Ji

= LWPRpredict (q, modeli)

Control using the learned model: ˙ q = ˆ J♯

1 ˙

ξ

⋆ 1 + ˆ

J2Pˆ

J1

♯ ˙

ξ

⋆ 2 − ˆ

J2ˆ J♯

1 ˙

ξ

⋆ 1

A babbling phase is necessary to

Initiate the models prior to any control → avoid completely random motions Cover roughly the whole learning space ← redundancy induces internal motions of the system and it is not sufficient to only learn a task along a specific trajectory

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 7 / 14

SLIDE 32

Learning strategy and learning tools

Learning strategy and tools

LWPR is an incremental function approximator which provides accurate approximation in very large spaces in O(k) Learning of a task Jacobian Ji: modeli = LWPRlearn (q, ξi) Prediction of the task Jacobian Ji: ˆ ξ1, ˆ Ji

= LWPRpredict (q, modeli)

Control using the learned model: ˙ q = ˆ J♯

1 ˙

ξ

⋆ 1 + ˆ

J2Pˆ

J1

♯ ˙

ξ

⋆ 2 − ˆ

J2ˆ J♯

1 ˙

ξ

⋆ 1

A babbling phase is necessary to

Initiate the models prior to any control → avoid completely random motions Cover roughly the whole learning space ← redundancy induces internal motions of the system and it is not sufficient to only learn a task along a specific trajectory

Learning is kept active while controlling in order to incrementally improve the models

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 7 / 14

SLIDE 33

Simulated experiments description

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 8 / 14

SLIDE 34

Simulated experiments description

◮ Under constrained case : positioning task (m1 = 2) of the end-effector

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 8 / 14

SLIDE 35

Simulated experiments description

◮ Under constrained case : positioning task (m1 = 2) of the end-effector ◮ Fully constrained case : positioning task (m1 = 2) of the end-effector and a one dimension positioning task for the elbow (m2 = 1)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 8 / 14

SLIDE 36

Simulated experiments description

◮ Under constrained case : positioning task (m1 = 2) of the end-effector ◮ Fully constrained case : positioning task (m1 = 2) of the end-effector and a one dimension positioning task for the elbow (m2 = 1) ◮ Over constrained case : positioning task (m1 = 2) of the end-effector and a positioning task for the elbow (m2 = 2)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 8 / 14

SLIDE 37

Results and Analysis

Under constrained experiment (with limited babbling : 2000 samples)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 9 / 14

SLIDE 38

Results and Analysis

Under constrained experiment (with limited babbling : 2000 samples)

֒ → A good babbling phase is required and on-line learning is necessary to incrementally improve the model.

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 9 / 14

SLIDE 39

Results and Analysis

Fully constrained experiment

˙ q = ˆ J♯

1 ˙

ξ

⋆ 1 + ˆ

J2Pˆ

J1

♯ ˙

ξ

⋆ 2 − ˆ

J2ˆ J♯

1 ˙

ξ

⋆ 1

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 10 / 14

SLIDE 40

Results and Analysis

Fully constrained experiment

˙ q = ˆ J♯

1 ˙

ξ

⋆ 1 + ˆ

J2Pˆ

J1

♯ ˙

ξ

⋆ 2 − ˆ

J2ˆ J♯

1 ˙

ξ

⋆ 1

֒

→ Error propagation due to the inversion of learned matrices is rather limited: the learned model is precise enough for this type of control closing the task space loop helps

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 10 / 14

SLIDE 41

Results and Analysis

Over constrained experiment

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 11 / 14

SLIDE 42

Results and Analysis

Over constrained experiment

֒ → The effectiveness of the method in terms of task hierarchy is illustrated: the secondary task has a limited impact on the achievement of the main task this secondary task is achieved with the minimum possible error

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 11 / 14

SLIDE 43

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 44

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 45

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 46

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment The dynamics mapping cannot be assumed as known and has to be learned

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 47

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment The dynamics mapping cannot be assumed as known and has to be learned ֒ → Salaun, C. and Padois, V. and Sigaud, O. “Learning inverse dynamics in a multiple tasks context”

. Submitted to ICRA 2010.

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 48

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment The dynamics mapping cannot be assumed as known and has to be learned ֒ → Salaun, C. and Padois, V. and Sigaud, O. “Learning inverse dynamics in a multiple tasks context”

. Submitted to ICRA 2010.

Experimentations are necessary on real robots involving more DoFs

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 49

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment The dynamics mapping cannot be assumed as known and has to be learned ֒ → Salaun, C. and Padois, V. and Sigaud, O. “Learning inverse dynamics in a multiple tasks context”

. Submitted to ICRA 2010.

Experimentations are necessary on real robots involving more DoFs ֒ → Clercq, C and Salaun, C. and Padois, V. and Sigaud, O. “On the Limitations of a Model Learning Approach for a Velocity

Controlled Humanoid Robot. ” Submitted to the IEEE Robotics and Automation Magazine.

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 50

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment The dynamics mapping cannot be assumed as known and has to be learned ֒ → Salaun, C. and Padois, V. and Sigaud, O. “Learning inverse dynamics in a multiple tasks context”

. Submitted to ICRA 2010.

Experimentations are necessary on real robots involving more DoFs ֒ → Clercq, C and Salaun, C. and Padois, V. and Sigaud, O. “On the Limitations of a Model Learning Approach for a Velocity

Controlled Humanoid Robot. ” Submitted to the IEEE Robotics and Automation Magazine.

Scaling up to higher number of DoFs raises active exploration questions which are not covered yet by our work.

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 51

Conclusions and Perspectives

This work demonstrates the effectiveness of a control approach combining model-based control in the task space and forward model-learning that could be use in order to cope with complexity and the need for adaptivity involved by new domains of Robotics However...

It becomes fully relevant to learn models when the robot is facing interaction with its environment The dynamics mapping cannot be assumed as known and has to be learned ֒ → Salaun, C. and Padois, V. and Sigaud, O. “Learning inverse dynamics in a multiple tasks context”

. Submitted to ICRA 2010.

Experimentations are necessary on real robots involving more DoFs ֒ → Clercq, C and Salaun, C. and Padois, V. and Sigaud, O. “On the Limitations of a Model Learning Approach for a Velocity

Controlled Humanoid Robot. ” Submitted to the IEEE Robotics and Automation Magazine.

Scaling up to higher number of DoFs raises active exploration questions which are not covered yet by our work.

The experimental work is developed with the iCub humanoid robot [15] (http://www.robotcub.org).

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 12 / 14

SLIDE 52

References

Willow Garage, “Overview of the PR2 robot,” http://www.willowgarage.com/pages/robots/pr2-overview.

Y. Nakamura, Advanced Robotics: redundancy and optimization.

Addison Wesley, 1991, ISBN 0-201-15198-7.

O. Khatib, “A unified approach to motion and force control of

robot manipulators: The operational space formulation,” IEEE Journal on Robotics and Automation, vol. 3, no. 1, pp. 43–53, Feb. 1987.

A. Ben Israel and T. Greville, Generalized Inverses : Theory and

Applications, 2nd ed. Springer, 2003, ISBN 0-387-00293-6.

K. Doty, C. Melchiorri, and C. Bonivento, “A theory of generalized

inverses applied to Robotics,” The International Journal of Robotics Research, vol. 12, no. 1, pp. 1–19, Feb. 1993.

A. Maciejewski and C. Klein, “Obstacle avoidance for kinematically

redundant manipulators in dynamically varying environments,” The International Journal of Robotics Research, vol. 4, no. 3, pp. 109–117, 1985.

S. Chiaverini, “Singularity-robust task-priority redundancy

resolution for real-time kinematic control of robot manipulators,” IEEE Transactions on Robotics and Automation, vol. 13, no. 3, pp. 398–410, June 1997.

N. Mansard and F. Chaumette, “Task sequencing for sensor-based

control,” IEEE Transactions on Robotics, vol. 23, no. 1, pp. 60–72,

Feb. 2007.
D. Nguyen-Tuong, J. Peters, M. Seeger, and B. Schoelkopf,

“Learning inverse dynamics: a comparison,” in Proceedings of the European Symposium on Artificial Neural Networks (ESANN), 2008.

J. Peters and S. Schaal, “Learning to control in operational space,”

The International Journal of Robotics Research, vol. 27, no. 2, pp. 197–212, 2008.

J. Baillieul, “Kinematic programming alternatives for redundant

manipulators,” in Proceedings of the International Conference on Robotics and Automation (ICRA), vol. 2, Mar. 1985, pp. 722–728.

A. D’Souza, S. Vijayakumar, and S. Schaal, “Learning inverse

kinematics,” in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), vol. 1, 2001, pp. 298–303.

L. Natale, F. Nori, G. Metta, and G. Sandini, “Learning precise 3d

reaching in a humanoid robot,” in Proceedings of the IEEE International Conference of Development and Learning (ICML), London, UK, July 2007.

S. Vijayakumar, A. D’Souza, and S. Schaal, “LWPR: A Scalable

Method for Incremental Online Learning in High Dimensions,” Edinburgh University Press, 2005.

G. Metta, G. Sandini, D. Vernon, L. Natale, and F. Nori, “The

iCub humanoid robot: an open platform for research in embodied cognition,” in PerMIS: Performance Metrics for Intelligent Systems Workshop, Washington DC, USA, Aug. 2008.

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 13 / 14

SLIDE 53

Control of redundant robots using learned models: an operational space control approach

Presented at the IEEE/RSJ International Conference on Intelligent Robots and Systems October 12, 2009 / St Louis - USA

by Camille Salaün, Vincent Padois (vincent.padois@upmc.fr), Olivier Sigaud

Université Pierre et Marie Curie Institut des Systèmes Intelligents et de Robotique (CNRS UMR 7222)

V. Padois (UPMC-ISIR)

Control of redundant robots using learned models ... IROS - 12/10/2009 - St Louis, USA 14 / 14