CUDA Particle SLAM

A high-performance particle filter implementation for Simultaneous Localization and Mapping (SLAM) and Monte Carlo Localization (MCL) using CUDA and exposed as a Node.js native addon.

⚠️ Linux Only: This project only supports Linux with NVIDIA GPUs. Windows is not supported.

Overview

This project implements a GPU-accelerated particle filter that can perform both:

• SLAM (Simultaneous Localization and Mapping): Build a map of an unknown environment while simultaneously tracking the robot's position
• MCL (Monte Carlo Localization): Localize a robot within a known map using particle filtering

The core algorithms run on CUDA-enabled GPUs for real-time performance with thousands of particles, while the Node.js interface makes it easy to integrate into robotics applications.

Features

• 🚀 GPU-Accelerated: All compute-intensive operations run on CUDA
• 🎯 Particle Filter: Supports up to 10,000+ particles for robust estimation
• 🗺️ Occupancy Grid Mapping: Uses Beta-Bernoulli model for probabilistic mapping
• 📊 Real-time Visualization: ASCII terminal visualizer for particle states and maps
• 📈 Performance Profiling: Built-in timing for all operations
• 🔄 Resampling: Efficient GPU-based particle resampling
• 📦 Easy Integration: Node.js API for seamless integration

Requirements

System Requirements

• Linux Operating System (Ubuntu 20.04+ recommended, WSL2 supported)
• NVIDIA GPU with CUDA support (Compute Capability 3.5+)
• CUDA Toolkit 10.0 or later (installed at /usr/local/cuda)
• Node.js 14.x or later recommended
• GCC 7+ with C++14 support

Dependencies

• node-addon-api: For Node.js native addon support
• node-gyp: For building native addons

Installation

1. Install CUDA Toolkit

Download and install the CUDA Toolkit from NVIDIA's website.

Ubuntu/Debian:

# Add NVIDIA package repositories
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/12.0.0/local_installers/cuda-repo-ubuntu2004-12-0-local_12.0.0-525.60.13-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-12-0-local_12.0.0-525.60.13-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2004-12-0-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cuda

# Verify CUDA installation
/usr/local/cuda/bin/nvcc --version

Ensure CUDA is installed at /usr/local/cuda. If installed elsewhere, create a symlink:

sudo ln -s /usr/local/cuda-12.0 /usr/local/cuda

2. Install Node.js Dependencies

npm install

3. Build the Native Addon

npx node-gyp configure
npx node-gyp build

Usage

SLAM (Mapping)

Build a map of an unknown environment:

# With visualization
node example_mapping.js

# Without visualization (faster)
node example_mapping.js --no-viz

This will:

1. Load sensor data from bigboi_munged.json
2. Run particle SLAM to build a map
3. Save the final map to node_baked_map.bin
4. Display profiling results

MCL (Localization)

Localize the robot in a known map:

# With visualization
node example_mcl.js

# Without visualization (faster)
node example_mcl.js --no-viz

This will:

1. Load a pre-built map from node_baked_map.bin
2. Load sensor data from one_loop_munged.json
3. Localize the robot using MCL
4. Display position estimates and profiling results

API Reference

ParticleSlam Class

const { ParticleSlam } = require('./index');

// Create instance
const slam = new ParticleSlam(
    num_particles,          // Number of particles (e.g., 5000-10000)
    max_trajectory_length,  // Maximum trajectory history per particle
    max_chunk_length,       // Maximum visual measurement chunks to buffer
    cell_size_m,           // Map cell size in meters (default: 0.1)
    pos_weight,            // Weight for positive observations (default: 0.7)
    neg_weight,            // Weight for negative observations (default: 0.4)
    alpha_prior,           // Alpha prior for Beta distribution (default: 1.0)
    beta_prior,            // Beta prior for Beta distribution (default: 1.5)
    measurement_saturation // Max observation counter value (default: 200)
);

Core Methods

Initialization

// Initialize CUDA memory and RNG
slam.init(random_seed = 1234);

Motion Update

// Apply dead reckoning motion step
slam.apply_step(
    { x: dx, y: dy, z: dtheta }, // Motion delta in robot frame
    timestamp,                     // Timestamp in seconds
    pos_std = 1.6e-3,             // Position noise std dev
    yaw_std = 1e-3                // Yaw noise std dev
);

Measurement Update

// Ingest visual measurement (60x60 occupancy grid)
const chunk = {
    timestamp: ts,
    cells: [[{ num_pos: 0, num_neg: 0 }, ...], ...]  // 60x60 array
};
const chunk_idx = slam.ingest_visual_measurement(chunk);

// For SLAM: accumulate map from particle trajectories
slam.accumulate_map_from_trajectories(chunk_idx);

// For MCL: accumulate map from global map
slam.accumulate_map_from_map(chunk_idx);

// Evaluate and resample particles
slam.evaluate_and_resample(chunk_idx);

State Estimation

// Calculate mean and covariance
const measurement = slam.calculate_measurement();
console.log(measurement.mean);        // { x, y, z }
console.log(measurement.covariance);  // { m: [9 elements] } (row-major 3x3)
console.log(measurement.is_gaussian); // boolean

Data Retrieval

// Download current particle states
const particles = slam.download_current_particle_states();
// Returns: [{ state: { x, y, z }, timestamp }, ...]

// Download particle scores
const scores = slam.download_scores();
// Returns: [score1, score2, ...]

// Download chunk states for a particle
const trajectory = slam.download_chunk_states_for_particle(
    particle_idx,
    num_chunks
);
// Returns: [{ x, y, z }, ...]

Map Operations

// For MCL: Set global map
slam.set_global_map(map);

// For MCL: Uniformly initialize particles across map
slam.uniform_initialize_particles();

// For MCL: Remove particles outside map bounds
slam.prune_particles_outside_map();

// Bake final map from best particle
const final_map = slam.bake_best_particle_map();

// For MCL: Bake global map from best particle
const updated_map = slam.bake_global_map_best_particle();

Utility Functions

// Load map from file
const { load_map_from_file } = require('./index');
const map = load_map_from_file('map.bin');

// Save map to file
const { save_map_to_file } = require('./index');
save_map_to_file(map, 'map.bin');

Map Structure

{
    width: 800,           // Map width in cells
    height: 600,          // Map height in cells
    min_x: -40.0,         // Minimum X coordinate (meters)
    min_y: -30.0,         // Minimum Y coordinate (meters)
    max_x: 40.0,          // Maximum X coordinate (meters)
    max_y: 30.0,          // Maximum Y coordinate (meters)
    cell_size: 0.1,       // Cell size in meters
    cells: [              // Flattened array of cells (height * width)
        { num_pos: 10, num_neg: 5 },
        // ...
    ]
}

Terminal Visualizer

The included ASCII terminal visualizer provides real-time feedback:

const { TermVisualizer } = require('./lib/term_visualizer');

const visualizer = new TermVisualizer({
    width: 120,           // Terminal width in characters
    height: 50,           // Terminal height in characters
    metersPerChar: 0.2,   // Meters per character cell
    maxParticles: 200     // Max particles to render
});

visualizer.init();
visualizer.setMap(map);  // Optional: set background map

visualizer.render(particles, mean, isGaussian, {
    title: 'SLAM Visualization',
    info: ['Position: (1.23, 4.56)', 'Particles: 5000']
});

Legend:

• ↑↗→↘↓↙←↖ - Blue arrows show particle positions and orientations
• ↑ (Green) - Mean pose when distribution is Gaussian
• ↑ (Yellow) - Mean pose when distribution is non-Gaussian
• █▓▒░ - Map occupancy (black = occupied, white = free)

File Structure

.
├── ParticleSlam.h          # Main particle filter header
├── ParticleSlam.cu         # CUDA implementation
├── src/
│   └── pswarm_wrapper.cpp  # Node.js addon wrapper
├── lib/
│   └── term_visualizer.js  # ASCII terminal visualizer
├── example_mapping.js      # SLAM example
├── example_mcl.js          # MCL example
├── binding.gyp             # Build configuration
├── index.js                # Node.js entry point
├── index.d.ts              # TypeScript definitions
└── README.md               # This file

Input Data Format

The examples expect JSON files with the following format:

[
  {
    "type": "dr_step",
    "ts": 1234567.89,
    "value": {
      "x": [x, y, theta]  // Robot state (odometry)
    }
  },
  {
    "type": "map_measurement",
    "ts": 1234568.12,
    "value": {
      "cells": [  // 60x60 grid
        [
          {"num_pos": 5, "num_neg": 2},
          null,  // No observation
          // ...
        ],
        // ...
      ]
    }
  }
]

Performance

Typical performance on an NVIDIA RTX 3080 with 10,000 particles:

Operation	Average Time
apply_step	0.5-1.0 ms
ingest_visual_measurement	0.3-0.5 ms
accumulate_map_from_trajectories	2-5 ms
evaluate_and_resample	3-8 ms
calculate_measurement	0.5-1.0 ms
download_current_particle_states	0.5-1.0 ms

Tips for Performance:

• Use --no-viz flag to disable visualization for maximum speed
• Adjust NUM_PARTICLES based on your GPU capability
• Larger max_trajectory_length increases memory usage
• Profile with built-in timing to identify bottlenecks

Troubleshooting

Build Issues

Error: nvcc: command not found

• Ensure CUDA Toolkit is installed at /usr/local/cuda
• Verify nvcc is accessible: /usr/local/cuda/bin/nvcc --version
• If CUDA is installed elsewhere, create a symlink:

  sudo ln -s /path/to/cuda /usr/local/cuda

Error: undefined symbol: _ZN6pswarm...

• The build system automatically sets rpath to /usr/local/cuda/lib64
• If issues persist, manually add to library path:

  export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

Error: Building on Windows

• Windows is not supported. Use Linux or WSL2 (Windows Subsystem for Linux 2)

Runtime Issues

Error: CUDA error: out of memory

• Reduce NUM_PARTICLES or MAX_TRAJECTORY_LENGTH
• Close other GPU-intensive applications

Visualization flickering

• The alternate screen buffer should eliminate flicker
• If issues persist, use --no-viz flag

Slow performance

• Profile with built-in timing to identify bottlenecks
• Ensure CUDA drivers are up to date
• Check GPU utilization with nvidia-smi

Algorithm Details

Particle Filter

Each particle maintains:

• Current state: (x, y, θ) in world frame
• Trajectory history of states at measurement times
• Individual occupancy map built from observations

Motion Model

Dead reckoning with Gaussian noise:

x' = x + Δx + noise(σ_pos)
y' = y + Δy + noise(σ_pos)
θ' = θ + Δθ + noise(σ_yaw)

Observation Model

60×60 occupancy grid centered on robot:

• Each cell contains positive/negative observation counts
• Beta-Bernoulli model for occupancy probability:
  • α = α_prior + pos_weight × num_pos
  • β = β_prior + neg_weight × num_neg
  • P(occupied) = α / (α + β)

Resampling

Systematic resampling based on particle scores:

• SLAM: Score based on map consistency
• MCL: Score based on match with global map

Contributing

Contributions are welcome! Areas for improvement:

• Additional sensor models (lidar, depth cameras)
• Loop closure detection
• Map serialization formats
• More visualization options

License

MIT License

Copyright (c) 2025

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Acknowledgments

This implementation uses CUDA for GPU acceleration and is designed for real-time robotics applications.