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.
Introduction
Collaborative robots (cobots) require precise spatial awareness to interact safely and effectively with their environment. While high-density LiDAR systems are standard for navigation, their bulk, weight, and cost often preclude their use as distributed sensors on a robot's end-effector or body. Miniaturized Time-of-Flight (ToF) sensors, such as the VL53L5CX, offer a compelling alternative. Weighing approximately 1 gram and costing under £7, these "tiny LiDARs" can be easily integrated into robotic manipulators.
However, these sensors present unique challenges: they provide extremely low-resolution data (typically an 8x8 grid) and are susceptible to noise influenced by distance and surface orientation. This paper addresses the gap in utilizing such "coarse" data for reliable object localization.
Sensor Characterization and Modeling
To transform noisy, low-resolution readings into reliable spatial data, we first conducted an extensive characterization of the VL53L5CX sensor using a Franka Panda robot arm.
Experimental Campaign
The sensor's performance was evaluated across a range of distances (20mm to 800mm) and orientations (up to 25°). Our findings revealed that the manufacturer's datasheet provides only a baseline; real-world performance exhibits significant offsets and systematic noise trends. Specifically, we observed a constant 0.6% standard deviation in the 100-800mm range, which deviates from reported specifications. Furthermore, orientation errors were found to follow a predictable "elliptic paraboloid" shape, representing a systematic bias rather than random noise.
Probabilistic Sensor Model (PSM)
Based on these observations, we developed a custom Probabilistic Sensor Model. We established a linear relationship between the true range ($z^*$) and the measured range ($\bar{z}$). Sensor noise was modeled as a Gaussian distribution where the variance is a function of the measured distance. This model allows for mathematical correction of orientation-induced biases using a second-order polynomial fit.
Object Localization via Particle Filter
We implemented a Monte Carlo Localization (MCL) framework using a Particle Filter (PF) to estimate an object's pose. The PF maintains a "belief" of the target's position and orientation, which is updated using the PSM.
By comparing real-time sensor beams against a virtual ray-casting of the object's CAD mesh, the filter converges on the most likely pose. This approach was tested on two targets: a geometrically simple crate and a complex "Oxford Deer" statue.
Results and Discussion
The proposed PSM consistently outperformed two baselines: an "Ideal Sensor" (IS) model assuming zero uncertainty and a "DataSheet" (DS) model based on manufacturer specs.
- Accuracy: For the crate, the PSM reduced position error by approximately 0.02m compared to baselines. Even for the complex deer statue, the system maintained a position error of < 0.035m and an orientation error of < 1.8°.
- Resolution and Geometry: A single 8x8 sensor proved sufficient for localizing simple geometric features. However, complex geometries like the deer statue highlighted the "coarseness" bottleneck, as fewer beams hit the target, reducing the impact of probabilistic corrections.
- Repeatability: The model demonstrated robustness across multiple physical sensor units, suggesting that a general model can be applied without per-unit calibration.
Conclusion
This work demonstrates that low-cost, miniaturized ToF sensors can enable effective spatial awareness for robotic manipulators when paired with an accurate probabilistic model. Future research will address the impact of material properties (reflectivity) and environmental factors like ambient lighting. Additionally, we aim to explore "active sensing" strategies to optimize viewpoint selection for faster and more accurate localization.
Summary updated