Follow-Ahead Robot — Quick Start

• Project Website: https://s-gemmin.github.io/follow-ahead-web/

Requirements

• ROS2 Humble
• QBot 2e with qbot_driver
• VICON system with vicon_ros2_node

Installation

Clone the repo into your ROS2 workspace src folder and build:

colcon build

Running on Hardware

Ensure ROS_DOMAIN_ID is set to the same value on both the QBot and your machine.

On the QBot:

# Terminal 1
python3 vicon_ros2_node_new.py

# Terminal 2
ros2 launch qbot_driver bringup.launch.py

On your machine:

# Terminal 1
ros2 run follow vicon_bridge

# Terminal 2
ros2 run follow main --ros-args -p sim:=false

---

Project Overview

Replication and ROS2 extension of the MCTS-DRL framework for proactive follow-ahead navigation on a physical QBot 2e. The system integrates Monte Carlo Tree Search with a Deep Reinforcement Learning value function and an LSTM-based human action predictor.

Key Features & Contributions

• ROS2 Architecture: Fully rebuilt and modularized from the original ROS1 codebase.
• Sim-to-Real Kinematics: Enforces strict hardware acceleration and velocity limits via fake_odom.py ($V_{\max}=0.6$ m/s, $a_{\max}=0.5$ m/s$^2$) to ensure simulations transfer to the QBot.
• Hardware Integration: Custom VICON Bridge (vicon_bridge.py and bringup_vicon.launch.py) for robust, ground-truth map-level EKF localization.
• Algorithmic Fixes: Corrected the reference MCTS implementation to prevent duplicate leaf node re-expansion and UCB corruption.
• Perception: Replaces default camera drivers with the official RealSense ROS2 wrapper, processing depth pointclouds via rtabmap_util for reliable obstacle detection.

---

Codebase & Workflow Guide

1. Training the RL Value Function

The A2C agent evaluates robot-human configurations to determine the "follow-ahead quality" for the MCTS planner. It is trained offline in an obstacle-free simulation.

cd src/follow-ahead-project/RL_sim
python3 train_a2c.py

2. Training the LSTM Predictor

The LSTM takes a rolling buffer of human poses and predicts the human's next action (straight, left, right) to bias the tree expansion.

cd src/follow-ahead-project/lstm-fc
python3 train_final_v3.py

3. Running Simulations

Test the algorithm in closed-loop simulation with kinematics constraints. RViz will open automatically.

# From workspace root (~/Desktop/qbot_ws)
source /opt/ros/humble/setup.bash && source install/setup.bash

# Run using ROS2 launch explicitly:
ros2 launch follow sim.launch.py test_case:=circle

(Available test cases: circle, stationary, square, oscillate, zigzag, gentle_arc, gentle_zigzag, approach_and_hold, straight. Kill with Ctrl+C to flush the data log to ~/follow_data/.)

4. Data Visualization

Process the flushed simulation logs to generate quantitative metrics and 2D trajectory plots.

python3 src/follow-ahead-project/scripts/plot_results.py ~/follow_data/ --summary