Scaling Agricultural Robotics with Cloud-Native ML Pipelines on AWS

Aigen's journey to scale its fleet of autonomous agricultural robots faced significant bottlenecks with its initial on-premises ML infrastructure. The core challenge was supporting a continuous model improvement cycle for hundreds of distributed edge robots, requiring efficient data ingestion from rural areas, high-throughput data labeling, and scalable model training. This led to a migration to a cloud-native architecture leveraging AWS services, particularly Amazon SageMaker AI, to build a robust MLOps pipeline.

Key Architectural Challenges Before Modernization

• Connectivity Constraints: Inconsistent internet access in rural farming areas complicated reliable data upload from robots to the cloud.
• High Data Labeling Cost: Manual labeling of thousands of images daily was expensive and time-consuming, hindering iteration speed.
• Limited Computational Power: On-premises GPUs (RTX 3090s) provided insufficient parallelism for specialized edge model training and fine-tuning large foundation models.
• Resource Contention: Model training and data labeling batch inference competed for the same limited on-premises GPU resources, leading to delays and inefficient workflows.

Cloud-Native Solution Architecture

Aigen adopted an AWS AI-driven, cloud-native approach to address these challenges, creating a closed-loop system for continuous model improvement. This architecture spans from data collection at the edge to iterative model training and rapid redeployment.

Model Architecture for Edge Computing

Aigen employs a hierarchical model architecture designed to balance accuracy with the stringent constraints of edge devices. This involves a progression from general-purpose Foundation Models to highly specialized Edge Models, optimizing for performance and resource usage at each stage:

• Foundation Models (L1): Proprietary and open-source vision models (e.g., SAM2, Grounding DINO) for general object recognition, segmentation, and synthetic data generation.
• Expert Models: Distilled from FMs, these perform precise, task-specific vision workloads, generating high-quality pre-labels for human review. They are larger (10s of millions of parameters) and use Vision Transformer and CNN architectures.
• Student Models: Compact, full-precision (FP32) models (<1.5M parameters) continuously fine-tuned on the latest data. Optimized for ultra-low latency and minimal memory usage through quantization-aware training (QAT) and pruning, achieving real-time performance on ~2 TOPS NPUs.
• Edge Models: Further optimized student models converted to TFLite with INT8 quantization for deployment on robot NPUs (1M-1.2M parameters, ~2MB memory, ~1.5W power consumption), sustaining double-digit FPS.

End-to-End MLOps Pipeline on AWS

1. Data Collection & Ingestion: Robots use AWS IoT Core to reliably offload raw data (video, telemetry, metadata) to Amazon S3 buckets, even in intermittent connectivity scenarios.
2. Data Processing & Labeling: An ETL pipeline preprocesses raw data. SageMaker AI processing jobs use an ensemble of expert models for automated pre-labeling. An active learning process then down-samples and prioritizes the most informative images for human-in-the-loop validation, significantly reducing manual effort and cost (22.5x cost reduction, 20x throughput increase).
3. Model Training: Final annotated data in Amazon S3 feeds SageMaker AI Training jobs. Multi-GPU instances (G5/G6 families) with Distributed Data Parallel (DDP) accelerate training of expert, student, and edge models. Edge-optimized models are deployed back to robots, while finetuned expert models improve the next cycle of automated labeling.