Particle Filters in AI & Artificial Intelligence

A Particle Filter represents the robot's position as a set of discrete Particles. Each particle is a 'Hypothesis'—it has a position (x, y) and an orientation (θ). At the start of a mission, if the robot is lost, we spread thousands of particles randomly across the map. This is called Global Localization. As the robot moves, every particle moves the same way. We are effectively simulating 1,000 different 'Parallel Worlds' to see which one matches reality.

When the robot's sensors see a landmark (like a wall 2 meters away), we Weight the particles. A particle that is 2 meters from a wall in its 'Parallel World' gets a high weight. A particle in the middle of an empty room gets a weight of zero. During the Resampling step, we randomly draw a new set of particles, but particles with high weights are much more likely to be picked (and cloned). Over several cycles, the cloud of particles 'Collapses' or Converges on the robot's true location.

Unlike Kalman Filters (which only track one 'Bell Curve'), Particle Filters are Non-Parametric. This means they can handle Multimodal Distributions. If a robot is in a building with two identical long hallways, the particles might form two distinct groups—one in each hallway. The robot effectively says, 'I am in either Hallway A or Hallway B.' Only when it sees a unique landmark (like a door) will one group of particles die out, allowing the robot to resolve the ambiguity. This robustness makes it the gold standard for indoor navigation.

What if your robot doesn't know where it is at all? Particle Filters use thousands of 'Digital Clones' (particles) to explore the map and find the most likely position of the real robot.

We start with a 'Cloud' of particles spread across the map. As the robot moves and senses, particles that don't match the sensor data 'Die' and are replaced by those that do.

Particle Filters are 'Non-Gaussian'. They can handle multiple guesses at once (like being in one of two identical hallways) until more data arrives.

Checkpoint: What happens to a particle that DOES NOT match the robot's sensor data?

• →It becomes the new leader
• →It is given a low weight and eventually removed during the Resampling step

This method is often called 'Monte Carlo Localization' (MCL). It is robust and used in everything from indoor vacuum robots to outdoor autonomous delivery bots.

By mastering Particle Filters, you learn to manage 'Global' uncertainty, allowing a robot to recover even if it is picked up and moved to a new location.

Checkpoint: Why is the 'Resampling' step important?

• →To save memory
• →To focus the computational power on the particles that are most likely to represent the robot's true position

Particle Filters mastered! You've learned to localize. Ready to explore 3D sensing with LiDAR and Radar?

Normalize Real Particle Weights. Finish normalizing particle weights so they sum to 1, ready to be treated as a probability distribution.