2025

Advancing 3D Object Detection for Safe and Efficient Autonomous Systems

Entrant

Category

Innovation in Technology - Best Autonomous Technology Innovation

Client's Name

Country / Region

United States

At Torc Robotics, we engineered a cutting-edge multi-modal perception system for autonomous vehicles, integrating LiDAR and camera sensor data with advanced geometric transformation and cross-attention aligners. This innovation significantly improved 3D object detection accuracy by 35% and boosted BEV semantic segmentation mIoU by 25%, addressing critical synchronization challenges across varying sensor timestamps. Leveraging a CenterNet-based detector and a high-performance DETR Transformer model, we enabled high-definition bird’s-eye-view (BEV) map generation, reducing lateral and longitudinal errors by 70% using only camera inputs.

We optimized real-time inference pipelines by converting PyTorch models to TensorRT, achieving inference speeds of 0.5 ms per frame and supporting multi-sensor processing at 10 fps. A custom ROS node was developed to handle diverse sensor inputs, ensuring low-latency detection and lane prediction with seamless integration into the vehicle’s decision-making system.

To ensure scalability and reliability, we deployed the system using Docker on AWS ECS clusters, achieving 100% uptime and robustness to multiple sensor failures in production environments. Asynchronous data loading and caching reduced memory usage by 40% and improved batch processing latency by 65%, enabling the system to process over 200,000 frames per second.

Additionally, we created a scalable dataset augmentation pipeline using Generative Adversarial Networks (GANs) and pre-trained AutoEncoder models, which improved object detection mAP by 19%. Our optimized Airflow pipeline for multi-sensor data preprocessing reduced the runtime from 25 hours to 8 hours, accelerating the model training process and enabling rapid iteration.

This multi-modal fusion framework is a game-changer for autonomous driving, delivering industry-leading performance, real-time scalability, and unmatched accuracy. With its robust architecture and advanced sensor fusion, this technology paves the way for safer, more reliable autonomous vehicles capable of operating seamlessly in complex environments.

Credits

 
2025
ASK JUNIOR Seaweed Fabric Series

Entrant

Hangzhou Ouchao Trading Co., Ltd.

Category

Innovation in Design - Fashion, Apparel and Garment Design

Country / Region

China

 
2025
Transforming Feature Deployment with Salesforce and LaunchDarkly Integration

Entrant

Ronith Pingili

Category

Innovation in Technology - Automation Technology

Country / Region

United States

 
2025
The Wind City Composer

Entrant

Cultural Affairs Bureau of Hsinchu City

Category

Innovation in Design - Conceptual Design

Country / Region

Taiwan

 
2025
Driving Digital Transformation: Prasad Manda’s Role in Cloud and ERP Modernization at 3M/Solventum

Entrant

Category

Innovation in Technology - Information Technology

Country / Region

United States