Safe and Interactive Crowd Navigation

This repository contains the code for our paper SICNav: Safe and Interactive Crowd Navigation using Model Predictive Control and Bilevel Optimization including the CrowdSimPlus Simulator.

Website | Repo | IEEE Xplore| arXiv | Video

@article{samavi2024sicnav,
  author={Samavi, Sepehr and Han, James R. and Shkurti, Florian and Schoellig, Angela P.},
  journal={IEEE Transactions on Robotics}, 
  title={SICNav: Safe and Interactive Crowd Navigation using Model Predictive Control and Bilevel Optimization}, 
  year={2024},
  volume={},
  number={},
  pages={1-19},  
  doi={10.1109/TRO.2024.3484634},
  url={https://arxiv.org/abs/2310.10982}}

Setup

In a conda environment with Python 3.8.13,

1. Install this package using pip:

   pip install -e .
   

2. Install the required libraries:

   pip install -r requirements.txt
   

3. Clone and install Python-RVO2 library. Note that a compatible version of Cython 0.29.33 will be installed already in step 2. You should only need to run:

   pip install -e <path-to-Python-RVO2-dir>/Python-RVO2/
   

4. (Recommended) Install HSL to use advanced solvers in IPOPT, the settings that we use in campc.py make use of a linear solver from HSL, however the code in this repo will work with the default IPOPT settings as well. Instructions: https://github.com/casadi/casadi/wiki/Obtaining-HSL.

Testing Crowd Navigation Algorithms

The following algorithms can be visualized by running the following command in the sicnav/ directory:

python simple_test.py --policy <policy>

• Collision Avoidance MPC (--policy campc) can be configured in sicnav/configs/policy.config to act as the following algorithms described in the paper:

  • SICNav-p (ours): our algorithm with access to privileged information about the humans
  • SICNav-np (ours): our algorithm without access to privileged information about the humans
  • MPC-CVMM: the MPC baseline that does not model human interactively

• Dynamic Window Approach (--policy dwa)

To train and visualize the Reinforcement Learning algorithms, SARL and RGL, please see the RL_nav/ subdirectory.

CrowdSimPlus Simulator

CrowdSimPlus is based on OpenAI gym and is an extension of CrowdSim (Repo, Paper), with the follwing added features:

• Static obstacles to allow for more realistic dense scenarios.
• Social Forces Model (SFM) policy to simulate human agents, in addition to the original Optimal Reciprocal Collision Avoidance (ORCA) policy. These policies can be found under crowd_sim_plus.envs.policy.
• Stable Baselines 3 (SB3) integration for Reinforcement Learning (RL) methods.

Contributors

This repository is primarily developed by Sepehr Samavi (sicnav and crowd_sim_plus) and James R. Han (RL_nav and crowd_sim_plus).