The Ground-Truth Loop
Follow the decisions that connect sensor recordings, reviewed labels, training data, and model evaluation.
A deployed model produces new evidence: missed objects, difficult grasps, unusual conditions, and successful behavior worth preserving. A useful data loop turns that evidence into a deliberate change to the next training set.
Avala connects the data work in that loop. Your training, evaluation, and deployment systems remain part of the workflow. Connecting them requires explicit integrations and quality gates.
The six stages
1. Ingest. Bring in supported sensor recordings, images, video, and point clouds. Preserve the information needed to interpret them, including schemas, timestamps, and calibration.
2. Visualize. Inspect the available sensor streams and compare their timing and geometry. Check the recording before treating it as annotation-ready.
3. Annotate. Use the labeling tools and supported auto-label models appropriate to the task. Review predictions and ambiguous cases with trained coworkers.
4. Curate. Choose the examples and quality states for a training set. Preserve the project configuration, source identifiers, and review context needed to understand that selection.
5. Train. Export reviewed data, validate it in your training loader, and record which export and code version produced the model.
6. Evaluate and deploy. Measure the model against relevant scenarios. Feed selected failures and new field data back into collection, labeling, or review.
Make the handoffs explicit
Moving data between tools does not establish that it is ready for training. Define what counts as an accepted label, who resolves disagreements, and which evaluation must pass before a model is deployed.
For example, a team may turn a difficult grasp into a new labeling task, review the contact-phase labels, add the accepted data to a release, and run a targeted evaluation. Those are workflow decisions. Automatic routing, retraining, and deployment happen only when the corresponding integrations and permissions are configured; there is no universal turnaround time.
Follow one decision through the loop
Keep this sequence in view:
recording → inspection → labels → quality review → dataset selection → training → evaluation → next data decision
At every handoff, ask what evidence the next person or system needs. A faster loop is useful when its outputs remain understandable and reproducible.
Next: Ingest & Visualize →