stonkens / cbf_opt

Toolbox for solving CBFs in Python

Control Barrier Function toolbox

Toolbox for implementing safety filters using Control Barrier Functions (CBFs) in python using cvxpy.

Control barrier functions are a principled tool to encode (through a scalar value function) and enforce (through a condition on the derivative of this value function) safety properties of a system. Enforcing safety using CBFs is typically implemented in a safety filter (also referred to as active safety invariance filter, ASIF) by minimally modifiying a nominal (safety-agnostic) policy to maintain safety. For control-affine systems, $\dot x = f(x) + g(x)u$, the CBF condition is linear in the control $u$ and can be efficiently solved online. This package provides an implementation of this using cvxpy. The solved constraint is then $u^* = \arg \min_u {\lVert u - u_{\text{nom}} \rVert_2^2 \mid \dot h + \alpha h \geq 0}$, with $h$ the CBF and $\alpha$ the maximal safety decay rate.

Installation

• Run pip install -e . to install this project and its dependencies (from requirements.txt).

Instructions

dynamics.py provides abstract classes for different types of dynamics, cbf.py for different type of CBFs, and asif.py for different types of safety filters. The toolbox is compatible with batched inputs (torch, tf or numpy) and individual inputs (numpy or jax)

Dependencies:

• cvxpy
• Soft dependency: experiment-wrapper: This package is experimental and not added to the requirements.txt file. Install it through pip install experiment-wrapper or downloading from source link To use this toolbox, a user defines the dynamics for their problem, the cbf, and the safety filter (solely requires setting $\alpha$ and the nominal policy). The user then can run experiments manually or using the experiment-wrapper package. Both use-cases are showcased in the examples folder.

An example for the control-affine setting, the Adaptive Cruise Control problem, is available in examples/acc.ipynb.