Glydways Deploys Pedestrian Detection via LiDAR Pipeline

Glydways, a company developing autonomous vehicles for urban mobility, has successfully deployed its first pedestrian detection model. To achieve this, the startup partnered with Uber AI Solutions to construct a scalable data pipeline specifically designed for Light Detection and Ranging, or LiDAR, technology. This infrastructure allowed Glydways to process complex 3D data efficiently, enabling the deployment of a critical safety feature in less than a quarter.

The primary challenge was not merely the volume of data, but the quality of the training inputs. LiDAR point clouds are difficult to interpret, especially at longer ranges where the density of points on a target object decreases. Glydways needed a rigorous operational process to ensure that every labeled frame contributed effectively to its perception models, avoiding the common pitfall of scaling data volume without maintaining precision.

Complexity of 3D Data Labeling

Unlike standard 2D image labeling, interpreting LiDAR data requires understanding spatial relationships in three dimensions. Farshid Moussavi, Head of Perception at Glydways, noted that this complexity is most acute when detecting objects far away, where the sensor returns fewer data points. Without a standardized approach, inconsistencies in labeling can lead to models that fail in real-world scenarios. The partnership focused on creating strict annotation standards and quality assurance workflows to mitigate this risk.

The teams collaborated to define clear criteria for what constitutes a valid pedestrian detection in a sparse point cloud. They implemented collaborative review processes to identify and correct labeling errors early. This structured approach ensured that the data fed into the model was consistent, which is essential for training a system that must operate safely in unpredictable urban environments.

Efficiency and Cost Reduction Outcomes

According to material reported by GN technics/ai (en-US), the collaboration yielded significant operational improvements. Glydways achieved 99% annotation accuracy within two months of starting the process. The team successfully labeled over 20,000 LiDAR frames, which provided a robust dataset for the initial model training. Furthermore, the streamlined pipeline resulted in an estimated 40% cost savings in the labeling phase compared to traditional methods.

These metrics indicate that the partnership balanced speed with precision effectively. By experimenting with different workflow structures, the teams identified the optimal balance between rapid data processing and high-quality output. This efficiency allowed Glydways to iterate on its models quickly, refining the pedestrian detector based on continuous feedback from the data pipeline.

Foundation for Future Expansion

The immediate result was a functional pedestrian detector, but the long-term value lies in the established operational capability. Glydways now possesses a repeatable pipeline for data ingestion, annotation, and quality review. This infrastructure allows the company to expand its perception capabilities into new object classes and more complex operating conditions without rebuilding the process from scratch.

Moussavi described the partnership as a force multiplier, extending the team’s capacity to deliver safe and reliable machine learning systems. The established pipeline serves as the operational foundation for Glydways’ broader goal of integrating autonomous vehicles into dedicated urban guideways. As the company scales its network, this data infrastructure will be critical for maintaining high standards of safety and reliability.