Glydways Builds Pedestrian Detection System with Uber Partnership

Autonomous vehicle startup Glydways has successfully developed its first pedestrian detection model by partnering with Uber AI Solutions. The collaboration focused on building a scalable pipeline for processing light detection and ranging, or LiDAR, data. This capability is critical for the company's goal of deploying autonomous shuttles that operate without stopping in dense urban environments.

The project resulted in the labeling of over 20,000 LiDAR frames with 99% annotation accuracy within two months. According to reporting from GN technics/ai (en-US), the partnership also delivered an estimated 40% reduction in labeling costs. The system was designed to handle the complex task of identifying pedestrians in three-dimensional point cloud data, a challenge that becomes significantly harder at longer ranges where data points are sparse.

Complexity of 3D point cloud labeling

LiDAR systems create a cloud of thousands of individual points to map the environment, rather than a continuous image. Identifying a human figure within this scattered data is inherently difficult. As distances increase, the number of points reflecting off a person decreases, making the shape less distinct. Glydways required a process that could consistently interpret these sparse patterns to ensure safety.

The company faced the trade-off between speed and precision. Early-stage training data was limited, meaning every error in the dataset could negatively impact the final model. The teams had to develop rigorous quality assurance workflows to balance rapid data ingestion with the strict accuracy required for safety-critical perception systems.

Operational standards for data quality

To solve this, the partners established specific annotation standards and collaborative review processes. Instead of relying solely on raw labeling volume, they focused on refining the operational workflow. This included experiments to find the optimal balance between efficiency and correctness, ensuring that the training data remained consistent as the perception models evolved.

The approach allowed Glydways to identify issues early in the process. By integrating quality checks directly into the data pipeline, the teams could maintain steady progress without compromising the integrity of the dataset. This structured method transformed data labeling from a simple task into a reliable engineering capability.

Scalability for future perception tasks

The primary outcome was not just a single dataset, but a repeatable infrastructure for future development. Glydways now possesses a structured pipeline for data ingestion, annotation, and quality review. This foundation allows the company to expand its perception capabilities to other object classes and more complex operating conditions without rebuilding the process from scratch.

This scalability is essential for Glydways' broader plans to integrate autonomous vehicles into urban flow networks. The ability to quickly generate high-quality training data for new scenarios reduces the time required to deploy new safety features. The partnership effectively created a force multiplier for the company's engineering team, enabling them to scale their perception models as their operational footprint grows.