Kodiak Deploys GigaFusionNet AI Model in 28 Driverless Trucks, Details Training Pipeline
Kodiak Robotics describes its GigaFusionNet foundation model and AI flywheel, which powers 28 driverless trucks, and outlines partnerships with Lambda and NVIDIA for training and deployment.
28
What Happened
Kodiak Robotics has revealed details of GigaFusionNet, its core AI foundation model for autonomous driving, which powers 28 driverless trucks operating with no humans in the cab as of March 31, 2026. The company presented its training approach at CVPR in Denver, emphasizing a multi-stage pipeline that includes data curation, pre-training, supervised fine-tuning, and an AI flywheel for continuous improvement.
28
Commercially operational with no humans in the cab as of March 31, 2026.
GigaFusionNet ingests multimodal sensor data from cameras, LiDAR, and radar to build a holistic understanding of the driving environment. The model is pre-trained on unlabeled data using self-supervised learning to learn general physical concepts, then specialized for tasks like 3D bounding-box detection and road geometry recognition.
Kodiak's AI flywheel uses autolabeling to generate training data from real-world miles, with a Teacher-Student model regime where a powerful teacher creates labels to train an efficient student. The company also partners with Lambda for NVIDIA HGX H100 accelerated computing infrastructure and with NVIDIA for model distillation onto the DRIVE Hyperion platform for in-vehicle deployment.
Why this matters
This explains how Kodiak uses self-supervised learning and autolabeling to scale autonomous trucking, demonstrating a path toward safe, commercially viable driverless vehicles.
Terms in This Story
- GigaFusionNet
- Kodiak's large-scale neural network foundation model that processes multimodal sensor data for autonomous driving.
- AI Flywheel
- A self-reinforcing loop where data from deployment continuously improves the AI model through autolabeling.
- Autolabeling
- A process where the AI automatically labels data without human intervention, enabling scalable training.
Summarised from the linked release; details can be imperfect — always verify against the original source.