Deploy & close the loop
The model ships, the fleet runs, and the failures it produces become the next dataset. This is the arrow that turns a pipeline into a flywheel.
A model in production is not the end of the loop — it's the most valuable part of the data source. Every deployment generates exactly the data you most need: the situations your current model handles badly. Capturing those and routing them back is what turns a one-way pipeline into a flywheel.
Watching the fleet
Once a model is deployed, the fleet starts producing recordings of its own. Avala's fleet tooling gives you the surface to manage that: a registry of devices with their status and firmware, the recordings they upload, events and markers that bookmark notable moments, recording rules that decide what's worth keeping, and alerts when something crosses a threshold.
The job at this stage is triage. Out of everything the fleet produces, which moments matter? A near-miss, an unexpected obstacle, a disengagement, a confidence drop — these are the frames worth pulling back into the loop. Events and markers let you flag them; rules let you capture them automatically.
Closing the loop
Here's where the whole course connects. A flagged failure from the field doesn't go into a ticket queue to be forgotten. It goes back to stage three:
deployed model fails → recording captured → routed to annotation → auto-labeled → human-verified through consensus → curated into the next dataset → retrained → redeployed.
Because annotation and verification are inside the platform rather than a separate vendor relationship, that round trip is measured in hours, not months. The anomaly you saw this morning can be verified ground truth in your training set this afternoon — with the lineage to prove how it was labeled and why.
This is the dynamic that made the leading autonomous-driving systems what they are: not algorithmic novelty alone, but billions of miles of edge-case data continuously improving the models. The moat is the loop, and the loop's speed is set by how fast you can turn a failure into trustworthy training data.
Why a closed loop beats a faster pipeline
It's tempting to optimize only for throughput — move more data, faster. But throughput on unverified data just trains your model on more noise. A closed loop optimizes for the thing that actually compounds: the rate at which real-world failures become correct labels. That's a different metric, and it's the one Avala is built to maximize.
The shape, one more time
You started with a raw recording and a robot that did something wrong. You ingested and visualized it in 4D, produced verified ground truth, curated a reproducible dataset, trained on it, deployed — and now the failures flow back to the start. The loop is closed.
The last lesson is about making this loop yours: the SDK, CLI, and agent integrations that let you wire every stage into your own systems.
Next: Make it yours →