Smart Wheelchair Autonomy

Dual 3D LiDAR autonomous wheelchair | ROS 2 + Nav2 + Cartographer SLAM | Hybrid A+DQN | 70% collision reduction | 80-120ms latency*

Overview

Autonomous navigation system for WHILL Model CR electric wheelchair using dual Velodyne VLP-16 3D LiDARs. Combines Google Cartographer SLAM for mapping, Nav2 for path planning, custom Hybrid A* planner, and a Dueling DQN for dynamic obstacle avoidance.

Research context: University of Delaware MS Robotics — assistive robotics for clinical environments (hospital autonomous wheelchair navigation).

Key Results

Metric	Value
Collision reduction (DQN vs baseline)	70%
End-to-end latency (P50)	88ms
End-to-end latency (P95)	132ms
SLAM map accuracy	< 5cm loop closure error
Navigation success rate	96.4% (250 trials)
Map coverage	3000+ m² (UD Medical Center)
Platform	NVIDIA Jetson AGX Xavier
Wheelchair	WHILL Model CR

System Architecture

2x VLP-16 LiDAR → PCL Merge → Cartographer SLAM (5cm map)
                    ↓
              EKF (wheel odom + IMU) → filtered pose
                    ↓
              Nav2 (Hybrid A* + Regulated Pure Pursuit)
                    ↓
              DQN Obstacle Layer (70% fewer collisions)
                    ↓
              WHILL CAN Bus Controller

Repository Structure

smart-wheelchair-autonomy/
├── src/
│   ├── dqn_obstacle_avoidance.py   # Dueling DQN controller
│   └── astar_planner.py            # Hybrid A* global planner
├── launch/
│   └── wheelchair_launch.py        # Full system ROS 2 launch
├── config/
│   ├── nav2_params.yaml            # Nav2 configuration
│   ├── cartographer/
│   │   └── wheelchair_2d.lua       # Cartographer SLAM config
│   └── ekf_params.yaml             # EKF state estimation
├── docs/
│   ├── architecture.md             # System architecture
│   ├── evaluation.md               # Performance metrics
│   └── proof_guide.md              # Evidence guide
├── results/
│   └── latency_measurements.md     # Detailed latency data
└── videos/
    └── README.md                   # Video demonstrations

Installation

# Install ROS 2 Humble
sudo apt install ros-humble-desktop ros-humble-nav2-bringup
sudo apt install ros-humble-cartographer-ros ros-humble-robot-localization
sudo apt install ros-humble-velodyne

# Clone and build
git clone https://github.com/DKrishna007/smart-wheelchair-autonomy.git
cd smart-wheelchair-autonomy
colcon build --packages-select smart_wheelchair_autonomy
source install/setup.bash

# Python dependencies
pip install torch>=2.0 numpy>=1.21.0

Quick Start

# Full autonomous navigation
ros2 launch smart_wheelchair_autonomy wheelchair_launch.py \
    slam_mode:=mapping dqn_enabled:=true rviz_enabled:=true

# Navigation only (with existing map)  
ros2 launch smart_wheelchair_autonomy wheelchair_launch.py \
    slam_mode:=localization map_file:=/path/to/map.yaml

# Simulation (Isaac Sim / Gazebo)
ros2 launch smart_wheelchair_autonomy sim_launch.py

DQN Training

# Train DQN in Isaac Sim
python src/dqn_obstacle_avoidance.py --train --episodes 5000

# Evaluate trained model
python src/dqn_obstacle_avoidance.py --eval --model models/dqn_wheelchair.pt

Hardware Configuration

Component	Model	Interface
Wheelchair	WHILL Model CR	CAN bus
LiDAR (×2)	Velodyne VLP-16	Ethernet
Compute	Jetson AGX Xavier	—
IMU	VectorNav VN-100	SPI
Battery	25.9V 31.2Ah	Internal

Citation

@misc{digamarthi2024wheelchair,
  title={Autonomous Navigation for Clinical Wheelchair using Dual LiDAR and DQN},
  author={Krishna Digamarthi},
  year={2024},
  institution={University of Delaware},
  url={https://github.com/DKrishna007/smart-wheelchair-autonomy}
}

Author

Krishna Digamarthi — MS Robotics, University of Delaware
GitHub: @DKrishna007