GithubHelp home page GithubHelp logo

arthursn138 / safe-pdp Goto Github PK

View Code? Open in Web Editor NEW

This project forked from wanxinjin/safe-pdp

0.0 0.0 0.0 25.77 MB

Safe Pontryagin Differentiable Programming (Safe PDP) is a new theoretical and algorithmic safe differentiable framework to solve a broad class of safety-critical learning and control tasks.

Home Page: https://wanxinjin.github.io/posts/safePDP

License: MIT License

Python 100.00%

safe-pdp's Introduction

Safe Pontryagin Differentiable Programming

This code project is an implementation of Safe Pontryagin Differentiable Programming (Safe PDP) methodology. Safe PDP is a new theoretical and algorithmic safe differentiable framework that can solve a broad class of safety-critical learning and control tasks---problems that require the guarantee of both immediate and long-term constraint satisfaction at any stage of the learning and control progress. Safe PDP has the following key features:

  • Safe PDP provides a systematic way for handling different types of constraints, including state and inputs (or mixed), immediate, and long-term constraints in a learning and control system.

  • Safe PDP attains the provable safety- and optimality- guarantees throughout the learning and control process, guaranteeing strict constraint satisification throughout the entire learning and control process.

  • Safe PDP establishes a unified differentiable programming framework to efficiently solve a board class of safety-critical learning and control tasks, such as safe policy optimization, safe motion planning, and safe learning from demonstrations, etc.

Please find the technical details of Safe PDP in our paper: https://arxiv.org/abs/2105.14937, by Wanxin Jin (Purdue University), Shaoshuai Mou (Purdue University), George J. Pappas (University of Pennsylvania)

This paper has been accepted in NeurIPS 2021.

1. Project Overview

The current version of the code project includes the following folders.

  • SafePDP: an independent package for core implementation of Safe PDP. This package includes two modules:

    • SafePDP.py the core module that implements the Safe PDP for constrained learning and control tasks
    • PDP.py the core module that implements the PDP (NeurIPS 2020, link) for unconstrained learning and control tasks
  • Examples: a folder containing different applications of Safe PDP, including

    • SPO: examples of applying Safe PDP for safe policy optimization.
    • LQR: examples of applying Safe PDP for safe motion planning.
    • MPC: examples of applying Safe PDP for
      • CODE: learning constrained ODEs from demonstrations
      • CIOC: jointly learning dynamics, state and input constraints, control cost from demonstrations.

    Note that all the above examples are presented in the Safe PDP paper. Please find the experiment details in the Safe PDP paper.

  • JinEnv: an independent package, containing a single module, JinEnv.py, which implements different control environments and its animation. More information about how to use JinEnv.py, please refer to link. Compared to the previous version of JinEnv.py, we here have added some interfaces to each environment to allow the user add the constraints to the environment.

  • ControlTools: an independent package, containing a single module ControlTools.py, which implements some popular control algorithms, including LQR, iLQR, system identification (DMD algorithms), ALTRO trajectory optimization, etc. Note that this package is irrelevant to Safe PDP, and It is included here mainly for convenience of users (who may use these algorithms for benchmark comparison).

2. Dependency

Note: before you try Safe PDP and JinEnv Packages, we strongly recommend you to familiarize yourself with the CasADi programming language, e.g., how to define a symbolic expression/function. Reading Sections 2, 3, 4 on the page https://web.casadi.org/docs/ is enough (around 30 mins)! Because this really helps you to debug and understand the codes here.

The codes have been tested and run smoothly with Python 3.7. on macOS Big Sur (11.2.3) machine.

3. How to Use the SafePDP Package

First, be relaxed :). We have optimized the interface of the Safe PDP package, which hopefully minimizes your burden to understand and use them. All variables/functions have pretty straightforward names, and most of the lines are carefully commented! In most of the cases, all you need to do is to specify/connect your control system from JinEnv.py environment to SafePDP.py, then Safe PDP will take care of the rest.

The quickest way to start is to read and run each example in the Example folder, e.g.,

  • Read and run Examples/SPO/SPO_Cartpole.py --- you will understand how to use SafePDP to solve safe policy optimization.

  • Read and run Examples/SPO/PO_Cartpole.py --- you will understand how to use SafePDP to solve unconstrained policy optimization.

  • Read and run Examples/SPlan/SPlan_Cartpole.py --- you will understand how to use SafePDP to solve safe motion planning.

  • Read and run Examples/Plan/Plan_Cartpole.py --- you will understand how to use SafePDP to solve unconstrained motion planning.

  • Read and run Examples/MPC/CIOC/CIOC_Cartpole.py --- you will understand how to use SafePDP to jointly learn dynamics, constraints, and control cost from demonstrations.

  • Read and run Examples/MPC/CODE/CODE_Cartpole.py --- you will understand how to use SafePDP to learning dynamics and constraints from demonstrations.

In SafePDP package, the instructions for the interfaces within PDP.py module can be found at here. The following briefly describes the interfaces within SafePDP.py module.

  • Class COCsys: this class has the following functionalities (interfaces):

    • construct a (parameterized) constrained optimal control agent, i.e., Equ. (1) in the Safe PDP paper,
    • solving a constrained optimal control problem by an OC solver,
    • differentiate the C-PMP and establish the auxiliary control system, i.e.,Theorem 1 in Safe PDP paper,
    • convert the constrained optimal control agent into an unconstrained one, i.e., Equ. (5) in Safe PDP paper,
    • Approximate both trajectory and its derivative using safe unconstrained couterparts, i.e., Theorem 2 in Safe PDP paper.
  • Class EQCLQR: this class is an implementation of an equality-constrained LQR solver, which is based on the paper of Efficient Computation of Feedback Control for Constrained Systems by F. Laine and C. Tomlin.

  • Class CSysOPT: this class particularly focuses on a special case of Safe PDP: for a constrained optimal control agent, the control cost and the constraints are given, only the agent dynamics model and control policy are parameterized (unknown). This class has the following functionalities (interfaces):

    • specify a (parameterized) dynamics model;
    • specify the control cost function;
    • specify the constraints;
    • specify the parametrized control policy: use the neural policy if you are solving the safe policy optimization, and use polynomial policy if you are solving the safe motion planning;
    • convert the constrained optimal control agent into an unconstrained one;
    • optimization via differential C-PMP.

Note that: all interfaces within SafePDP.py module are corresponding to the technical details of the Safe PDP paper. Please read the Safe PDP paper carefully before understanding these interfaces.

4. Citation

If you have difficulty to understand the code, please contact me by email: [email protected]

If you find this work helpful in your research, please consider citing our paper

  @article{jin2021safe,
    title={Safe pontryagin differentiable programming},
    author={Jin, Wanxin and Mou, Shaoshuai and Pappas, George J},
    journal={Advances in Neural Information Processing Systems},
    volume={34},
    pages={16034--16050},
    year={2021}
  }

safe-pdp's People

Contributors

wanxinjin avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.