Tiny LiDARs for Manipulator Self-Awareness: Sensor Characterization and Initial Localization Experiments

Giammarco Caroleo, Alessandro Albini, Daniele De Martini, Timothy D. Barfoot, Perla Maiolino

This paper, recently published by researchers from the Oxford Robotics Institute and the University of Toronto, explores how "tiny LiDARs"—specifically low-cost, miniaturized Time-of-Flight (ToF) sensors—can provide robots with the spatial self-awareness needed for manipulation and inspection tasks.

TL;DR

By systematically characterizing the noise and error patterns of $£6.50$ miniaturized ToF sensors, researchers developed a probabilistic model that enables precise object localization. Using a Particle Filter, they demonstrated that even "coarse" 8x8 depth maps can accurately estimate the pose of complex objects, outperforming standard manufacturer datasheet models.

What Problem Does This Paper Address?

For collaborative robots to work safely and effectively, they need to know where objects are in their workspace. While traditional LiDAR systems provide high-density point clouds, they are often too bulky, heavy, or expensive for distributed use on a robot's body or end-effector.

Miniaturized ToF sensors (like the VL53L5CX) are an attractive alternative: they weigh only 1 gram and cost very little. However, they provide extremely low-resolution data (typically just an 8x8 grid) and are prone to noise influenced by distance and surface orientation. Most existing localization algorithms (like SLAM) are designed for high-density data and struggle with this "coarseness." The authors address the gap of how to turn these noisy, low-res "tiny LiDARs" into reliable localization tools.

Methods

The researchers' approach follows three main phases:

  1. Sensor Characterization: Using a Franka Panda robot arm, they conducted an extensive experimental campaign to see how the sensor's readings changed based on distance (20mm to 800mm) and orientation (up to 25°).
  2. Probabilistic Sensor Model (PSM): They moved beyond the manufacturer's datasheet to create a custom model. They found that the true range ($z^*$) relates linearly to the measured range ($\bar{z}$) and that orientation errors follow a predictable "elliptic paraboloid" shape. They modeled the sensor noise as a Gaussian distribution where the variance is a function of the measured distance.
  3. Monte Carlo Localization (MCL): They implemented a Particle Filter (PF) that uses this PSM to update the "belief" of an object's pose. By comparing real-time sensor beams against a virtual ray-casting of the object's mesh, the filter converges on the most likely position and orientation of the target.

Datasets

The study utilized custom-collected data rather than standard public benchmarks:

  • Characterization Data: 100 samples per pose across various distances and angles to establish the baseline error curves.
  • Localization Targets:
    • A Crate: Representing simple geometric features (planes and corners).
  • A Deer Statue: Representing complex, irregular geometry. The ground truth for this was derived from the "Oxford Deer Point Cloud."
  • Baselines: The PSM was compared against a "DataSheet" (DS) model and an "Ideal Sensor" (IS) model that assumes zero uncertainty.

Key Insights and Ideas

  • Datasheets are just a starting point: The researchers found significant offsets and noise trends in real-world use that the manufacturer's datasheet didn't capture. Specifically, they found a constant 0.6% standard deviation in the 100-800mm range, which differed from the reported specs.
  • Orientation is a predictable bias: The error introduced by the angle of the sensor relative to the target surface isn't just random noise—it's a systematic bias that can be mathematically corrected using a second-order polynomial fit.
  • Resolution isn't everything: One of the most surprising "aha" moments is that a single 8x8 sensor is often sufficient to localize a simple object like a crate. Adding more sensors or more data samples provided diminishing returns for simple geometries.
  • Geometry Sensitivity: For complex objects like the deer statue, the "coarseness" of the sensor becomes a bottleneck because fewer beams hit the target, making the probabilistic corrections less impactful compared to simpler shapes.

Key Results

  • Superior Accuracy: The proposed PSM consistently outperformed both the Datasheet and Ideal Sensor models.
  • Crate Localization: The PSM achieved a position error approximately 0.02m lower than the baselines on average.
  • Complex Objects: Even with the irregular deer statue, the system maintained a position error of < 0.035m and an orientation error of < 1.8°.
  • Robustness: The model proved repeatable across three different physical sensor units, suggesting that a single general model can be used without needing to calibrate every individual sensor off the assembly line.

Limitations and Future Work

The authors acknowledge several areas for further exploration:

  • Material Properties: The current study used only white vinyl. Different colors, textures, and materials (which affect reflectivity) will likely alter the ToF readings and require further characterization.
  • Environmental Factors: The experiments were conducted under controlled lighting and temperature; real-world industrial environments might introduce more interference.
  • Active Sensing: The current work uses offline data samples. Future work could involve "active" localization, where the robot moves its arm specifically to gain the most informative viewpoints of an object (Next-Best-View planning).
  • Multipath Interference: ToF sensors often struggle with "light bouncing" in corners; accounting for this could further improve accuracy in cluttered scenes.

Summary updated

Abstract

For several tasks, ranging from manipulation to inspection, it is beneficial for robots to localize a target object in their surroundings. In this paper, we propose an approach that utilizes coarse point clouds obtained from miniaturized VL53L5CX Time-of-Flight (ToF) sensors (tiny lidars) to localize a target object in the robot's workspace. We first conduct an experimental campaign to calibrate the dependency of sensor readings on relative range and orientation to targets. We then propose a probabilistic sensor model that is validated in an object pose estimation task using a Particle Filter (PF). The results show that the proposed sensor model improves the performance of the localization of the target object with respect to two baselines: one that assumes measurements are free from uncertainty and one in which the confidence is provided by the sensor datasheet.
BibTeX
@inproceedings{caroleo2025tiny,
  author = {Giammarco Caroleo and Alessandro Albini and Daniele De Martini and Timothy D. Barfoot and Perla Maiolino},
  title = {{Tiny LiDARs for Manipulator Self-Awareness: Sensor Characterization and Initial Localization Experiments}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11246575},
}

Loading related content...

Figure