Leader-Follower Navigation with Social Force Model Integration

This repository contains the implementation of a Leader-Follower navigation algorithm integrated with the Social Force Model (SFM) for socially-aware robot navigation in crowded environments. We compare our approach against an MPC-based baseline across multiple challenging scenarios.

---

Table of Contents

• Overview
• Project Structure
• Installation
• Usage
• Evaluation Scenarios
• Evaluation Metrics
• Reproducing Results
• Citation
• License

---

Overview

Navigating robots through dense human crowds requires balancing efficiency with social compliance. This project introduces a Leader-Follower approach where the robot identifies and follows pedestrians heading toward similar destinations, leveraging human navigation expertise to achieve socially compliant paths.

Key Contributions:

• A Leader-Follower planner that dynamically selects pedestrian leaders based on goal alignment and trajectory prediction
• Integration with Social Force Model for reactive collision avoidance
• Comparative evaluation against Model Predictive Control (MPC) baseline
• Novel evaluation metrics including Path Smoothness and Invasion of Personal Space

---

Project Structure

ROB599LeaderFollowerAlgorithm/
├── simulator.py              # Base simulation environment with evaluation metrics
├── mpc_controller.py         # Social Force compliant MPC controller
├── leader_follower.py        # Leader-Follower planner implementation
├── LeaderFollowerSFM.py      # Leader-Follower + SFM integration module
├── environment.yml           # Conda environment specification
├── evals/                    # Evaluation notebooks for all scenarios
│   ├── BlackFridayLF.ipynb   # Black Friday scenario - Leader-Follower
│   ├── BlackFridayMPC.ipynb  # Black Friday scenario - MPC
│   ├── ConventionLF.ipynb    # Convention scenario - Leader-Follower
│   ├── ConventionMPC.ipynb   # Convention scenario - MPC
│   ├── JunctionLF.ipynb      # Junction scenario - Leader-Follower
│   ├── JunctionMPC.ipynb     # Junction scenario - MPC
│   ├── MuseumLF.ipynb        # Museum scenario - Leader-Follower
│   ├── MuseumMPC.ipynb       # Museum scenario - MPC
│   ├── PromenadeLF.ipynb     # Promenade scenario - Leader-Follower
│   ├── PromenadeMPC.ipynb    # Promenade scenario - MPC
│   ├── RoundaboutLF.ipynb    # Roundabout scenario - Leader-Follower
│   ├── RoundaboutMPC.ipynb   # Roundabout scenario - MPC
│   └── analyze_results.py    # Results analysis script
├── blender/                  # 3D visualization files
│   ├── BlackFriday.blend     # Black Friday scenario visualization
│   ├── Convention.blend      # Convention scenario visualization
│   ├── Junction.blend        # Junction scenario visualization
│   ├── Museum.blend          # Museum scenario visualization
│   └── Roundabout.blend      # Roundabout scenario visualization
├── results/                  # Evaluation logs (JSON format)
├── mpc_params_search/        # MPC parameter grid search notebooks
└── media/                    # Generated visualizations

Module Descriptions

Module	Description
simulator.py	Core crowd simulation environment with agents, obstacles, Social Force Model, and evaluation metrics (path smoothness, personal space invasion)
mpc_controller.py	Model Predictive Control local planner with social force-aware cost functions for goal reaching, obstacle avoidance, and personal space preservation
leader_follower.py	Leader-Follower planner that identifies optimal pedestrian leaders based on goal alignment, velocity matching, and position scoring
LeaderFollowerSFM.py	Integration layer combining Leader-Follower subgoal planning with SFM for reactive navigation

Blender Visualizations

The blender/ directory contains 3D visualization files for each evaluation scenario. These .blend files can be opened with Blender (version 3.0+) to visualize the simulation environments and generate high-quality renders of the navigation scenarios.

File	Scenario
BlackFriday.blend	High-density shopping environment
Convention.blend	Convention hall with corridors
Junction.blend	Intersection crossing
Museum.blend	Museum gallery space
Roundabout.blend	Circular pedestrian area

---

Installation

Prerequisites

• Miniconda or Anaconda
• Python 3.10+

Setup Environment

1. Clone the repository:

   git clone https://github.com/sebhelge/ROB599LeaderFollowerAlgorithm.git
   cd ROB599LeaderFollowerAlgorithm

2. Create the conda environment:

   conda env create -f environment.yml

3. Activate the environment:

   conda activate comphri_hw5

4. Verify installation:

   python -c "import numpy; import matplotlib; print('Installation successful!')"

---

Usage

Running the Simulator

from simulator import CrowdSimulator, Agent, SocialForceModel

# Initialize simulation
sim = CrowdSimulator(time_step=0.25, max_steps=200)

# Add robot and humans
sim.add_robot(x=0, y=0, gx=10, gy=10)
sim.add_human(x=5, y=5, gx=0, gy=0)

# Set navigation policy
from LeaderFollowerSFM import Leader_Follower_SFM
policy = Leader_Follower_SFM(sim=sim)
sim.set_robot_policy(policy)

# Run simulation
while not sim.done:
    sim.step()

# Calculate metrics
metrics = sim.calculate_metrics()

Using the MPC Controller

from mpc_controller import MPCLocalPlanner

planner = MPCLocalPlanner(
    simulation=sim,
    horizon=10,
    dt=0.25,
    max_speed=1.4,
    wg=1.0,   # Goal weight
    ws=1.0,   # Static obstacle weight
    wd=1.0,   # Dynamic obstacle weight
    wps=0.5   # Personal space weight
)

---

Evaluation Scenarios

We evaluate our approach across six challenging crowd scenarios:

Scenario	Description
Black Friday	High-density chaotic shopping environment with erratic pedestrian movements
Convention	Dense bidirectional pedestrian flow in narrow corridors
Junction	Multi-directional crossing at an intersection
Museum	Sparse crowds with unpredictable pedestrian behavior
Promenade	Linear walkway with mixed pedestrian flow patterns
Roundabout	Circular pedestrian flow with continuous merging and diverging

---

Evaluation Metrics

Our evaluation framework includes the following metrics:

Metric	Description
Time to Goal	Number of simulation steps to reach the goal
Completion Rate	Percentage of runs where the robot successfully reaches the goal
Path Efficiency	Ratio of actual path length to straight-line distance
Average Speed	Mean velocity throughout the trajectory
Path Smoothness	Measures trajectory roughness based on angular deviations
Personal Space Invasion	Count of instances where robot enters human personal space (3m bubble)
Minimum Human Distance	Closest approach to any pedestrian during navigation
Collision Rate	Percentage of runs resulting in collision

---

Reproducing Results

Step 1: Activate Environment

conda activate comphri_hw5

Step 2: Navigate to Evaluation Directory

cd evals/

Step 3: Run Scenario Notebooks

Open and execute the Jupyter notebooks for each scenario:

Scenario	Leader-Follower	MPC Baseline
Black Friday	BlackFridayLF.ipynb	BlackFridayMPC.ipynb
Convention	ConventionLF.ipynb	ConventionMPC.ipynb
Junction	JunctionLF.ipynb	JunctionMPC.ipynb
Museum	MuseumLF.ipynb	MuseumMPC.ipynb
Promenade	PromenadeLF.ipynb	PromenadeMPC.ipynb
Roundabout	RoundaboutLF.ipynb	RoundaboutMPC.ipynb

jupyter notebook ConventionLF.ipynb

Step 4: Analyze Results

python analyze_results.py ../results/convention_LeaderFollower_n100_20251207_210535.json

Pre-computed Results

All evaluation results are stored in the results/ directory as JSON files, containing metrics from 100 randomized runs per scenario.

---

Citation

This project implements the Leader-Follower approach from the following paper:

@misc{liao2025followingneedrobotcrowd,
      title={Following Is All You Need: Robot Crowd Navigation Using People As Planners}, 
      author={Yuwen Liao and Xinhang Xu and Ruofei Bai and Yizhuo Yang and Muqing Cao and Shenghai Yuan and Lihua Xie},
      year={2025},
      eprint={2504.10828},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2504.10828}, 
}

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Acknowledgments

• ROB 599: Human-Robot Interaction, University of Michigan