Introducing "Stages": Precision Over-The-Air Model Deployment for Your Edge Fleets
As developers scale their spatial intelligence applications on the Willow Dynamics platform, model libraries grow fast. When you are tracking hundreds of unique human motions, gestures, and actions, a new operational challenge emerges:
How do you ensure the right edge devices get exactly the right AI models - and nothing else?
Today, we are thrilled to announce Stages.
Stages is our new deployment management layer for the Cloud Oracle. It allows you to group specific .int8 Relative Distance Matrix (RDM) models into targeted "Stages" (like deployment environments or payload groups). Your edge applications simply poll their assigned Stage API endpoint to dynamically sync only the models they need.
It’s not rocket science, but it is necessary for scaling spatial intelligence across enterprise fleets, unlocking seamless Over-The-Air (OTA) updates and eliminating multi-account headaches.
Here are just a few ways our partners can use Stages to build cooler, smarter applications:
The Multi-App Studio (Unified Accounts)
Many of our enterprise partners build fundamentally different products. You might have a smart-yoga app tracking poses, and a commercial warehouse tracking system monitoring lifting ergonomics. Previously, developers had to spin up separate Willow accounts to keep the apps from downloading each other's models. Now, you simply deploy the yoga models to Stage: Yoga App Prod and the lifting models to Stage: Warehouse Tracker, managing your entire ecosystem under one unified billing and dashboard account.
Dynamic Fitness Content "Packs"
Imagine a digital fitness platform that wants to release a "Push-Up Mastery" update. Instead of pushing a massive, monolithic app update through the iOS/Android app stores, developers can instantly OTA deploy a group of 15 new models (Diamond, Wide-Grip, Explosive, Pike push-ups) directly to the Stage: Advanced Fitness Tier. Next time the user's app boots up, it seamlessly pulls the new models.
State-Based Logic in Robotics
Robots don't need to load every action model into RAM at all times. A robotic system can swap its active Stage based on its current physical environment. For example, a robotic arm transitions from a "Sorting Phase" to an "Assembly Phase". Upon transitioning, its logic controller pings the Cloud Oracle to drop the Stage: Sorting Actions and instantly load the Stage: Assembly Actions, keeping RAM usage incredibly lean and DTW matching lightning-fast.
Multi-Angle & Omnidirectional Grouping
Sometimes a single action looks vastly different depending on the camera placement. For smart-security or cashier-less retail environments, engineers can group different topological views of the exact same action (e.g., "Item Placed in Pocket" from Top-Down, Profile, and Face-On angles) into a single Stage: Checkout Cam Payload. The edge device downloads the group, running them concurrently to ensure 360-degree recognition accuracy.
Heterogeneous Edge Fleets
If you are deploying across a mix of hardware - say, an XR headset running Qualcomm chips, a security camera running an Ambarella NPU, and an industrial IoT sensor from a different manufacturer - they may have different compute thresholds. You can use Stages to partition "High-Fidelity Models" for the XR headsets, and "Lite Models" for the lower-power edge nodes, ensuring optimal performance across disparate hardware.
How to use it
Stages is live in your dashboard right now. Navigate to the Manage -> Stages tab to create your first deployment group.
For our integration engineers, your edge SDKs can simply call: GET /api/v1/data/stages/{stage_id}
The Oracle will instantly return the stage metadata along with an array of short-lived, pre-signed S3 URLs for every .int8 binary model assigned to that stage, ready for immediate, secure download to the edge.