Precision farming powered by K3s and TensorRT
summary
Aurea Imaging builds a TreeScout: a rugged Jetson edge box running Python microservices orchestrated with K3s on Jetson, converting PyTorch YOLO models to TensorRT to cut inference from ~22 ms to ~6.7 ms per frame, and mapping tree-level blossom while the tractor moves at 8 km/hr driving speed. Deployed to 40 devices, it produces hourly actionable maps for blossom-thinning, fixed many field failures with containerized rollbacks, and still needs work on scalability and OS updates.
AI-written summary · about
chapters (20)
- 0:00Introduction to precision farming
- 1:47Becoming an apple grower in 5 minutes
- 3:56Blossom thinning use case
- 6:46Initial challenges with tree detection
- 8:51Rethinking the tech stack
- 10:14Edge device software struggles
- 12:23Choosing Kubernetes over ROS
- 13:49Modern stack on the edge
- 15:53Success in 2024 season
- 16:36Blossom thinning in action
- 17:40Bug fixes with OTA updates
- 18:23Lessons learned from the journey
- 19:51Scaling limitations in cloud backend
- 20:11TensorRT performance boost
- 21:58MLOps for model accuracy
- 22:20NVIDIA Jetson Python challenges
- 23:47Future of over-the-air updates
- 24:49Expanding hardware and analytics
- 25:10Scaling fleet to 100 devices
- 25:32Q&A session
related talks
55mLarge Scale Distributed AI on the JVM
32mKubernetes and AI To Protect Our Forests: A Cloud Native Infrastructure for Wildfire Prevention
45mExtracting real-time data at the edge with SageMaker & NVIDIA GPUs
30mAI On-board of Satellites, Towards Autonomous Scientific Instruments
34mHow to Train Your Dragon and Its Friends: AI on the Edge with Eclipse Kura
10mReady, Set, Go for K8s at edge today for AI/ML with Tensorflow
28mGeoAI: Evolution of AI-Based Tree/Vegetation Detection Models Using Imagery
26mBuilding a Cloud Native Edge Platform for Accelerating AI Inference