Glydways Builds Autonomous Walking Detection System
A new autonomous vehicle startup has successfully deployed a pedestrian detection model by creating a specialized data pipeline with Uber, achieving high accuracy in under three months.
Autonomous mobility company Glydways has successfully developed and deployed its first pedestrian detection model, a critical step for operating its new category of urban transport. The company partnered with Uber AI Solutions to build a scalable data pipeline that processed light detection and ranging data, allowing the system to reliably identify people in real-world environments. This achievement was reached in less than a quarter, marking a significant milestone for the startup's ability to move riders safely without stopping.
The core challenge was not just labeling images, but creating a consistent operational process for handling complex 3D sensor data. Glydways builds 'Flow Networks' that integrate purpose-built vehicles with dedicated guideways, a setup that requires the software to distinguish pedestrians from other objects with extreme precision. According to Farshid Moussavi, Head of Perception at Glydways, the company could not launch its services without this specific capability, making the quality of the training data a matter of safety rather than just technical performance.
Complexity of three-dimensional sensor data
Unlike standard cameras, LiDAR systems generate dense clouds of points to map the physical world. This data is notoriously difficult to label because the number of points representing a person decreases significantly with distance, making detection harder at longer ranges. Glydways needed more than just volume; it required a workflow that could maintain high accuracy while the perception models evolved. The partnership focused on developing specific annotation standards and quality assurance workflows to handle this technical complexity, ensuring that every labeled frame contributed effectively to the model's learning.
Efficiency gains in annotation workflow
The collaboration resulted in labeling over 20,000 LiDAR frames with an estimated 40 percent cost savings compared to traditional methods. By experimenting with different workflow structures, the teams found a balance between speed and precision that allowed for continuous refinement of labeling quality. This approach also provided strategic recommendations on data collection, strengthening the overall training dataset. For a company with limited early-stage data, this efficiency was crucial, as every annotation had to be correct to maximize the value of the limited resources available.
Foundation for future perception tasks
The immediate result was a deployed pedestrian detector, but the long-term value lies in the established operational capability. Glydways now possesses a structured pipeline for data ingestion, annotation, and quality review that can be reused for new scenarios. This repeatable process allows the company to expand its perception capabilities to additional object classes and more complex operating conditions without starting from scratch. As reported by GN technics/ai (en-US), this infrastructure serves as a force multiplier for the engineering team, ensuring that future safety features can be developed with the same level of rigor and speed.
