Vibly: viability in state-action space

A Python framework for working with viable sets in state-action space, learning safety constraints, and for evaluating the effect of morphology on robustness and stability. This code accompanies the following papers:

1. A Learnable Safety Measure Steve Heim, Alexander von Rohr, Sebastian Trimpe, Alexander Badri-Spröwitz, Conference on Robot Learning 2019. Paper-specific examples in demos/measure_learning/. Code examples

2. Beyond Basins of Attraction: Quantifying Robustness of Natural Dynamics Steve Heim and Alexander Badri-Spröwitz IEEE Transaction on Robotics 2019. Note: This code implements the core algorithm for the paper, but does not reproduce all results and figures. The Arxiv version contains fewer typos and is recommended.

3. A Little Damping Goes a Long Way: a simulation study of how damping influences task-level stability in running Steve Heim, Matthew Millard, Charlotte Le Mouel, Alexander Badri-Spröwitz Royal Society Biology Letters 2020. Paper-specific examples in demos/damping_study/. Code examples

Installation

We recommend using a virtual environment. Note, GPy is only required for the safe learning examples, and can be safely removed from the requirements. Install from your terminal with

conda (recommended):
conda create -n vibly python=3.9 conda activate vibly then follow the same instructions as for pip: pip install -r requirements.txt && pip install -e .

PyEnv + PipEnv:
pipenv install -r requirements.txt
pipenv install -e .

pip:
pip install -r requirements.txt
pip install -e .

You should now be able to run all examples in demos/, and import packages as following:

# import brute-force viability tools
import viability as vibly
# import a model of a dynamical systems
from models import slip
# import classes for active sampling and learning
from measure import active_sampling

Use for viability

Examples are shown in the demos folder. We recommend starting with slip_demo.py, and then computeQ_slip.py.
The viability package contains:

• compute_Q_map: a utility to compute a gridded transition map for N-dimensional systems. Note, this can be computationally intensives (it is essentially brute-forcing an N-dimensional problem). It typically works reasonably well for up to ~4 dimensions.
• parcompute_Q_map: same as above, but parallelized. You typically want to use this, unless running a debugger.
• compute_QV: computes the viability kernel and viable set to within conservative discrete approximation, using the grid generated by compute_Q_map.
• get_feasibility_mask: this can be used to exclude parts of the grid which are infeasible (i.e. are not physically meaningful)
• project_Q2S: Apply an operator (default is an orthogonal projection) from state-action space to state space. Used to compute measures.
• map_S2Q: maps values of each state to state-action space. Used for mapping measures from state space to state-action space.

Reproduce CoRL safe learning study

You will need to first regenerate the ground-truth data used for comparison, by running demos/computeQ_hovership.py and demos/computeQ_slip.py.
Then, select and run an experiment by running run_learning_examples.py in demos/measure_learning/. The experiment details, including algorithm hyper-parameters and initialization, are defined in the experiment file, e.g. demos/measure_learning/hovership_default.py.
The learning algorithm is split between measure/active_sampling, which includes sampling strategy, and measure/estimate_measure, which handles everything dealing with the measure in different spaces. If you're looking to use the measure for a different learning approach, you probably want to look into the active_sampling.py file.

Reproduce RSBL damping study

The code to reproduce the results from the paper are in /demos/damping_study/. You can run compute_measure_damping.py, which will generate all the data needed; however, this can take a long time (~20 hours on a 24-core desktop). If you just want to inspect the results, all the pre-computed data (and code) can be downloaded from Dryad. We encourage you to use this code, which may have improvements/bugfixes, and simply copy/paste the dataset from data/guineafowl into the data folder.

Create your own dynamics

You can easily add your own models in the models package. The viability code does expect a few helper functions which need to be implemented as well. We recommend using the hovership.py example as a template, as it is simple and heavily commented. For examples of simulating a more complex system, see the slip.py model.

Contact

Feel free to open an issue, or get in touch via [email protected] for any questions, comments, etc.