3 ways auto-labeling in Viam can elevate your edge ML project

To train computer vision models for edge deployment, you need quality labeled data. And as your fleet of edge devices grows and data collection accelerates, it’s crucial to label those datasets efficiently. Annotating images has been a manual (and tedious) process for developers—until now. I’m excited to share how Viam’s new auto-labeling feature transforms this essential process into a streamlined, scalable workflow that gets your models to production faster. 

Auto-labeling transforms existing machine learning models into intelligent labeling assistants, and dramatically accelerates the process from data collection to deployed model. This lets developers focus on building a great model while automation handles the repetitive annotation work. Here are three key ways Viam’s auto-labeling feature can elevate your next edge ML project.

1. Spend less time labeling data

Auto-labeling uses trained models to annotate new images automatically. Instead of manually labeling every image in a large set, developers can use a already-trained model to generate predictions for new images. Developers can then review and adjust the annotations, and use the new dataset to train another model. This approach reduces labeling time from hours to minutes, which means faster iteration and improved model performance.

Easily auto-label images in a dataset using an existing model

Imagine you’re building a quality inspection system for a robotic pizza production line. You manually labeled 200 images of common quality issues—uneven toppings, burnt edges, and incorrect portion sizes. With auto-labeling, you can use this initial model to annotate the next 2,000 images captured by a camera attached to the production conveyor, and simply verify that the annotations are accurate.

Instead of spending days drawing bounding boxes around misplaced pepperonis or uneven cheese distribution, auto-labeling lets you review and fine-tune predictions in under an hour. With Viam, you can deploy your pizza inspection system days or weeks earlier, and ensure that only perfectly crafted pizzas are delivered to your customers.

This isn’t just about speed—it’s about consistency and accuracy. Models apply the same labeling logic across your entire dataset, eliminating the variability that creeps in when humans label images over multiple sessions, or when different team members interpret labeling guidelines differently. When you spot an error pattern, you can quickly reject those predictions and try again. The result? Higher quality training data in a fraction of the time.

2. Merge data for continuous training 

Developers can also save time by using a continuous training workflow—which Viam supports by merging auto-labeled datasets with existing ones. This means less time waiting to amass a large enough data set to train the initial model, and more opportunities to iterate along the way. 

Suppose you’re developing an obstacle detection model for a food delivery robot. Your initial model, trained on 500 images, handles well-lit streets just fine, but struggles with dark, shadowy conditions. As your robot makes deliveries, it captures new images that are auto-labeled using your existing model. You quickly review and correct any mislabeled obstacles and navigable paths, merge this new data with your original dataset, and retrain the model.

Within a week, your robot can confidently navigate at dusk and recognize partially obscured curbs—something that would have taken months if you waited to collect and manually label thousands of images before retraining.

Continuous training allows models to improve steadily as new data is captured: each cycle strengthens your model’s performance while your dataset grows organically. Your model learns from edge cases it encounters in production and can adapt to real-world conditions. It’s the difference between waiting months to deploy a “perfect” model and quickly getting a working model into production that can improve over time. 

3. Manage your data—and everything else—in one place

Because auto-labeling happens directly on the Viam platform, instead of via a third-party tool, your data stays in one place and is accessible to your entire team in the cloud. This greatly reduces complexity and increases efficiency in ML workflows: data flows seamlessly from collection to deployment, developers don’t have to juggle multiple tools, manage API keys, or wrestle with data export/import workflows, and larger teams can collaborate effectively. 

Quickly review predictions by accepting (A) or rejecting (R) labels

Picture that you’re building a custom object detection system to track warehouse inventory. Your cameras capture images of packages throughout the day, and automatically sync to the Viam cloud. When you have 100 new images, you build a new dataset and use auto-labeling with your latest model to identify package types and damage indicators. You review these labels while sipping your morning coffee, merge the verified data with your training set, and start the training process for a new model—all from one platform. 

By lunchtime, the improved model deploys to all your warehouse cameras without exporting data, switching between platforms, or losing context. Your entire ML pipeline—from camera to deployed model—lives in one place, making it simple to manage what used to be complicated workflows.

Viam’s unified platform also fosters flexibility that transforms how teams can respond to real-world challenges. When machines encounter new scenarios in the field, the models can quickly adapt by using new captured data; retraining models on auto-labeled datasets; and deploying improvements across entire fleets without switching platforms or losing context.

Developers can quickly iterate and improve while maintaining full visibility into what’s driving the model’s decisions by tracing them from inference back to the training data. Rather than being locked into rigid workflows across disconnected tools, teams can pivot quickly as requirements change, troubleshoot issues faster, and scale confidently—all because everything works together seamlessly within Viam.

Accelerate your ML workflow

Auto-labeling isn’t just convenient—it’s the most sustainable way to build ML datasets at scale. Combined with our merge datasets capability, it enables a continuous training workflow that keeps your models improving as your data grows. Whether you’re building quality inspection systems, autonomous navigation, or custom object detection, the labeling bottleneck no longer has to slow you down.

If you’re tired of the manual labeling grind, it’s time to let your models do the heavy lifting. Start with a small labeled dataset, train your first model, and watch as auto-labeling accelerates your journey to production-ready ML on the edge.