Lab 11: Localization on the Real Robot

Objective

The goal of this lab was to implement localization on a real robot using the Bayes filter, focusing exclusively on the update step based on 360-degree ToF sensor scans. Because the robot's motion was too noisy for accurate odometry, the prediction step was omitted. This lab emphasized the challenges of translating algorithms from simulation to real-world systems, and made use of an optimized localization framework to enable real-time performance on the robot.

Implementation

I modified the Arduino code from the Mapping Lab to collect exactly 18 ToF distance measurements—fewer than the continuous stream used previously—and send them over Bluetooth. When the robot receives the PERFORM_OBSERVATION_LOOP command, it initiates a rotation and transmits the sensor readings once the loop is complete.

Initially, I encountered issues with invalid yaw data after making large changes to the previous working code. To address this, I adopted an iterative debugging approach: making one small change at a time, testing to confirm the robot still returned the expected data, and repeating until I reached the desired result. A custom notification handler was written to process incoming Bluetooth data.

Both the data collection and reception logic were encapsulated within the PERFORM_OBSERVATION_LOOP command, which I called each time I wanted the robot to collect and transmit a full sensor sweep.

Code:

Code snippet 1 Code snippet 2 Code snippet 3 Code snippet 4

Results

The robot experienced mechanical difficulties during turning—specifically, one wheel often struggled to spin—resulting in uneven rotation and significant drift. This made consistent, centered rotation difficult and introduced error into the localization estimates.

The following poses were tested, and each belief estimate reflects my best attempt at achieving a clean, stable rotation at each location. 

(-3 ft, -2 ft, 0°)
(0 ft,  3 ft, 0°) 
(5 ft, -3 ft, 0°) 
(5 ft,  3 ft, 0°)

Results:

Result at (-3, -2) Result at (0, 3) Result at (5, -3) Result at (5, 3)

Discussion

Compared to the simulation-based Bayes filter from Lab 10, implementing localization on the real robot posed several additional challenges. The primary issue was hardware reliability—specifically, motor inconsistencies that caused the robot to drift or rotate unevenly. This mechanical drift introduced error into the sensor sweep and likely caused a mismatch between the real-world environment and expected map views.

Despite this, the Bayes filter was still able to generate belief estimates that roughly corresponded to the correct regions of the map. In more constrained or uniquely shaped areas of the environment, the robot localized more accurately. However, in open regions or near symmetric features, the localization contained more error.


Back to Main Page