top of page

Optimizing Navigation Transformers for AAM

Writer's picture: Mehrnaz SabetMehrnaz Sabet

Updated: Dec 10, 2024

This Fall, we embarked on a transformative journey to push the boundaries of AI optimization for autonomous navigation. Our project envisions a future where AI seamlessly powers drones navigating complex urban airspaces. A pivotal milestone in this journey was our participation in the exclusive NASA Open Hackathon, a collaborative effort between OpenACC, NASA, and Nvidia, held as part of the Open Hackathons Program.





Our selection for this hackathon was no small feat, as participants had to undergo a rigorous application process to secure their spot. This exclusive event brought together top innovators and provided unparalleled access to cutting-edge resources, including Nvidia DGX Cloud and dedicated mentors. Over the course of one month, we worked intensively to optimize training pipelines for navigation transformers tailored to domain-specific use cases in Advanced Air Mobility.


During the hackathon, we focused on accelerating training for two Vision-Language-Action (VLA) models, specifically pretraining and finetuning them for drone navigation using both custom simulation-generated and real-world navigation data. While the task and model were smaller in scale compared to Orion’s broader goals, this experience was invaluable in several ways:


  • Developed efficient training pipelines optimized for training large-scale models that will be our project's main building blocks.

  • Gained critical insights into compute needs of our project using DGX cluster nodes.


Our work during the hackathon yielded impressive results:

  • A smaller, quantized navigation model that maintained comparable performance while being 40% more memory efficient.

  • Increased training performance by 55% using techniques like quantization, mixed precision, and distributed training.


These pipelines are now being reused to efficiently train models for navigation layers of Orion, ensuring minimal resource waste and reduced time-to-deployment. Beyond the hackathon, we also developed strategies that balance the trade-off between generalizability and domain-specific performance for Advanced Air Mobility.


Sharing Our Work: I am thrilled to announce that I will be presenting our work at the Open Accelerated Computing  (OAC) Summit. My talk, Accelerating AI for Autonomous Navigation: Optimizing Navigation Transformers for Large-Scale Use Cases, will cover:

  • The methodologies we developed during the hackathon and beyond.

  • The challenges we overcame in optimizing AI for large-scale domain-specific use cases.

  • How these efforts are shaping the future of Advanced Air Mobility.


Participating in this collaborative hackathon reinforced the value of partnerships and community-driven innovation. The summit is a perfect platform to share our work, inspire others, and explore new frontiers in accelerated AI.


Join Us: The summit will feature leading voices in accelerated computing and AI, making it an unmissable event for anyone passionate about the future of AI and its applications. I look forward to engaging with the community and showcasing how collaborative innovation can drive real-world impact. Read more below:





22 views

Recent Posts

See All
bottom of page