Dynamical movement primitives: Learning attractor models for motor behaviors
read more
Citations
An Algorithmic Perspective on Imitation Learning
Imitation Learning: A Survey of Learning Methods
Recent Advances in Robot Learning from Demonstration
A tutorial on task-parameterized movement learning and retrieval
Learning Physical Collaborative Robot Behaviors From Human Demonstrations
References
Pattern Recognition and Machine Learning
Pattern Recognition and Machine Learning
Applied Nonlinear Control
Pattern Recognition and Machine Learning (Information Science and Statistics)
Real-time obstacle avoidance for manipulators and mobile robots
Related Papers (5)
Learning Stable Nonlinear Dynamical Systems With Gaussian Mixture Models
Frequently Asked Questions (13)
Q2. What have the authors contributed in "Dynamical movement primitives: learning attractor models for motor behaviors" ?
In this paper, the authors propose a generic modeling approach to generate multidimensional systems of weakly nonlinear differential equations.
Q3. What is the useful property of modeling behaviors in a dynamical systems framework?
A useful property of modeling behaviors in a dynamical systems framework comes from the scaling properties and invariance properties that can be designed into dynamical systems.
Q4. How can one influence the temporal evolution of their dynamical systems without affecting the transformation system?
By modulating the canonical system, one can influence the temporal evolution of their dynamical systems without affectingthe spatial pattern generated by the transformation system.
Q5. What is the definition of a nonlinear forcing term?
The nonlinear forcing term can be represented as an autonomous coupling term that can be learned with standard machine learning techniques that are linear in the open parameters.
Q6. What is the coupling term for obstacle avoidance?
the coupling term adds a movement perpendicular to the current movement direction as a function of the distance vector to the obstacle (see Hoffmann et al., 2009, for more details).
Q7. How did the authors use the model to generate movement to six different targets?
Then the authors applied the model to generate movement to six different targets, distributed with 60 degrees difference on a circle around the origin.
Q8. How many DOFs did the robot need to perform?
these tasks required the coordination and phase locking of 30 DOFs, which was easily and naturally accomplished in their approach.
Q9. What is the motivation to present in this letter?
The large variety of follow-up and related approaches to their initial work on dynamical movement primitives is one of the motivations to present in this letter the theory, insights, and a refined approach to learnable dynamical systems that the authors hope will continue to attract even more active research in the future.
Q10. What are the main differences with their approach?
The main differences with their approach is that the underlying dynamics is much more complex than ours (with several hundreds of state variables), that reservoir computing does not offer proof of stability of learned attractors, and that it is less easy to incorporate feedback terms for online trajectory modulation.
Q11. What is the way to model the dynamical systems?
From a practical point of view, one should first carefully investigate what properties a model requires in terms of temporal and spatial invariance and then realize these properties by choosing the most appropriate variant of the dynamical systems model and the most appropriate coordinate system for modeling.
Q12. What is the definition of a nonlinear dynamical system?
The explicit time dependence of this nonlinearity, however, creates a nonautonomous dynamical system or, in the current formulation, more precisely a linear time-variant dynamical system.
Q13. What could be used to exclude such cases?
Such cases could be excluded by more sophisticated classifiers that would employ, for example, confidence levels in decision making.