One Hell of a Lyapunov Learner.

This code implements Learning Control Lyapunov Functions, or CLFs, by leveraging the Stable Estimation of Dynamic Systems paper from Aude Billard's group at EPFL. For details, please consult

   S.M. Khansari-Zadeh and A. Billard (2014), "Learning Control Lyapunov Function
   to Ensure Stability of Dynamical System-based Robot Reaching Motions."
   Robotics and Autonomous Systems, vol. 62, num 6, p. 752-765.

One may find the MATLAB version of this implementation in the matlab subfolder.

Khansari-Zadeh has a subtle example that illustrates the advantages of SEDS over DMPs, LWPRs, regress_gauss_mixs etc in his 2014 Autonomous Systems paper, and reproduced below:

Learning Stable Task-Space Trajectories.

For recorded WAM robot end-effector trajectories on a 2D plane (there are two pre-recorded demos available in the data directory), the task is to stabilize pre-recorded trajectories using a combo of GMMs, Gaussian Mixture Regression, and Control Lyapunov Functions -- all learned from data.

Learning Stable Trajectories

S-shaped demos and reconstructions from three different initial conditions. The left image below denotes a demonstration of the robot drawing the letter S from three different initial conditions, and converging to an attractor at the origin; while the right image denotes the Gaussian Mixture Regression-based CLF that corrects the trajectories in a piecewise manner as we feed the algorithm the data.

W-shaped demos from three different initial conditions.

Setup.

Dependencies: Scipy | Numpy | Matplotlib.

Install as follows:

  pip install -r requirements.txt .

And that about wraps up setting stuff up!

Usage

Basic Python Usage:

  python scripts/demo.py

Python Usage [with options]:

  python scripts/demo.py [--silent|-si] [--model|-md] <s|w>  [--pause|-pz] <1e-4> [--visualize|-vz] [--kappa0|-kp] <.1> [--rho0|-rh] <1.0> [--traj_num|-tn] <20e3>

where angle brackets denote defaults.

Options

• --silent/-si: Optimize the control Lyapunov function in silent mode.

• --model/-md: Which saved model to use? w or s.

• --visualize/-vz: Do we want to visualize the regions of attractor (ROA) of the Lyapunov dynamics?

• --off_priors/-op: Do we want to use the offline computed priors with KZ's Expectimax algo implementation?

• --pause_time/-pz: Time between updating the stabilization of the dynamical system's reaching motions on the pyplot display screen.

• --kappa0/-kp: Exponential coefficient in the class-Kappa function that guarantees V is positive outside the equilibrium point/region (see paper referenced above for details).

• --traj_num/-tn: Length of time to simulate trajectory corrections.

• --rho0/-rh: Coefficient of class-Kappa function (see paper referenced above for details).

Example python usage [with options]:

  python scripts\demo.py -pz .1 --silent --model s -tn 20000                                               

Jupyter Notebook Interactive Example

Please find the example in the file clf_demos.ipynb.

FAQS

• Why Gaussian and Weighted Regression and not Neural Networks

Gaussian mixtures and locally-weighted regression are efficient and accurate, as opposed to neural networks which are computationally intensive and approximate at best. For real-world autonomous systems, we cannot afford any under-approximation of our system dynamics; otherwise, we'll enjoy the complimentary ride that travels on the superhighway of disgrace to hell.

• Why Consider this CLF correction scheme for stabilizing trajectories over statistical learning methods or dynamic movement primitives?

  • Dynamic Movement Primitives are do not do well when it comes to learning multiple demos;

  • Statistical Learning approaches, on the other hand, really do not have a guaranteed way of ensuring the learned dynamics are Lyapunov stable;

• What is different between this and the matlab implementation?

  • Well, for starters, a cleaner implementation of the Gaussian mixture models/regressors used in estimating dynamics along every trajectory sample.

  • A straightforward CLF learner.

  • A straightforward nonlinear optimization of the CLF cost that handles both inequality and equality constraints.

  • Written in python and easy to port to other open-source robot libraries.

Methods

• Through a clever re-parameterization of robot trajectories, by a so-called weighted sum of asymmetric quadratic functions (WSAQF), and nonlinear optimization, we learn stable Lyapunov attractors for the dynamics of a robot's reaching motion, such that we are guaranteed to settle to non-spurious and stable attractors after optimization;

• This code leverages a control Lyapunov function in deriving the control laws used to stabilize spurious regions of attractors, and eradicate limit cycles that e.g. expectation-maximization may generate;

• This code is pretty much easy to follow and adapt for any dynamical system. Matter-of-factly, I used it in learning the dynamics of the Torobo 7-DOF arm in Tokyo ca Summer 2018.

TODOs

• Add quivers to Lyapunov Function's level sets plot.

• Add options for plotting different level sets for the Lyapunov Function.

• Intelligently initialize the Lyapunov Function so optimization iterations do not take such a long time to converge.

• Fix bug in WSAQF when L is started above 1 for a demo.

Citation

If you use LyapunovLearner in your work, please cite it:

@misc{LyapunovLearner,
  author = {Ogunmolu, Olalekan and Thompson, Rachel Skye and Pérez-Dattari, Rodrigo},
  title = {{Learning Control Lyapunov Functions}},
  year = {2021},
  howpublished = {\url{https://github.com/lakehanne/LyapunovLearner}},
  note = {Accessed October 20, 2021}
}