Glydways and Uber Build Autonomous Pedestrian Safety Pipeline

Glydways, a company developing a new category of urban mobility known as Flow Networks, has completed its first pedestrian detection model. The system relies on a specialized data pipeline built in partnership with Uber AI Solutions to ensure the autonomous vehicles can safely navigate city streets. This collaboration was essential because the company could not launch its service without a mechanism to reliably identify people in its environment.

The partnership focused on processing LiDAR data, a technology that maps the physical world using light pulses. For autonomous vehicles, this data is the primary source of information about obstacles and pedestrians. According to reporting by GN technics/ai (en-US), the project achieved 99% annotation accuracy within two months, labeling over 20,000 frames while reducing labeling costs by an estimated 40%. This efficiency allowed Glydways to move from concept to a deployed detector in less than a quarter.

Why Pedestrian Detection Is Critical

The core challenge for Glydways is that its vehicles operate in dense urban environments where pedestrians are unpredictable. Farshid Moussavi, the Head of Perception at Glydways, stated that the company cannot deploy its system without reliably detecting pedestrians throughout its entire network. Unlike traditional cars that may stop frequently, Flow Networks are designed to move riders on-demand without stopping, which raises the stakes for safety systems. A failure to detect a person crossing a dedicated guideway could have severe consequences, making this specific capability a prerequisite for any launch.

The technical difficulty lies in the nature of LiDAR data. As Moussavi explained, creating 3D labels from point clouds is complex, particularly at longer ranges where the density of data points on a target decreases. In these scenarios, a human or an object may appear as a sparse collection of dots rather than a solid shape. This sparsity makes it difficult for algorithms to distinguish a pedestrian from background noise or other static objects, requiring a level of precision that standard data labeling processes often fail to provide.

Building a Scalable Data Operation

Rather than attempting to build this labeling capability from scratch, Glydways partnered with Uber AI Solutions to establish a scalable operation. The teams worked together to develop specific annotation standards and quality assurance workflows. This process involved collaborative reviews to continuously refine labeling quality while improving efficiency. The goal was not just to label data, but to create an operational process capable of producing consistently high-quality training data as the perception models evolved.

The collaboration included experimenting with different workflows to find the optimal balance between speed and precision. Recommendations regarding data collection and annotation strategy further strengthened the training dataset. Moussavi noted that open communication and teamwork were key to identifying issues early. He described the experience as the smoothest he has had with a data labeling partner, citing Uber’s quality, responsiveness, and prior experience with 3D labeling at scale as compelling factors in the decision to partner.

Creating a Repeatable Safety Foundation

The result of this engagement was more than a high-quality dataset; it was the establishment of a repeatable operational capability. Uber AI Solutions enabled Glydways to obtain its first externally labeled dataset for pedestrians, which directly led to the first internally trained and deployed pedestrian detector. This structured pipeline for data ingestion, annotation, and quality review provides a foundation that Glydways can continue to use as it expands into new perception scenarios.

With a proven data pipeline in place, the company can now expand its perception capabilities to include additional object classes and tasks. This includes handling increasingly complex operating conditions that may arise as the network grows. Moussavi summarized the value of the collaboration by stating that the most valuable result was a reliable labeling pipeline and partnership that the company can count on for future tasks. He viewed Uber AI Solutions as a force multiplier that extends the team’s capacity to deliver a safe and reliable machine-learning-based system.