Fleet learning
Turn every failure in your fleet into the data your model needs next.
Avala runs the loop around your model. We ingest your rollouts, find where the model fails, capture the data that covers the gap and score the next release on real-world scenes it has never seen. You train the model, own it and decide when it ships.
Fleet Learning Infrastructure for Physical AI
The loop
Seven stages, and one record for each cycle.
Each cycle starts with what your fleet did and ends with a release you choose to ship. Avala runs five of the stages. Training and the release decision stay with you.
Observe
Run by Avala
Your robots record rollouts. Avala ingests the videos and logs with their device, recording and event context.
Diagnose
Run by Avala
Avala segments each rollout, clusters the failures into a failure taxonomy and maps each cluster to coverage cells.
Evaluate
Run by Avala
The current model is scored on held-out real-world scenes, so the starting point is measured before anything changes.
Acquire
Run by Avala
A collection brief becomes a capture campaign: which tasks, which homes or workplaces, which objects and how many sequences.
Train
Run by you
You train on the new dataset release with your own code and compute. Avala keeps the lineage from sequence to release.
Validate
Run by Avala
Avala scores the new model on the same held-out cohort and reports both versions side by side, with confidence intervals.
Release
Run by you
You decide whether the model ships. The decision and its evidence go into the record for the next cycle.
The record
A capability experiment ties each cycle together.
Every cycle is written down as one capability experiment. It names what the team wants to improve and keeps the evidence in one place, so any result traces back to the data and the decision behind it.
- Model version
- The model the fleet ran when the failure appeared.
- Failure condition
- The behavior that went wrong, named in the failure taxonomy.
- Hypothesis
- What the team expects new data to change, written down before collection starts.
- Coverage cells
- The tasks, settings and objects the new data must cover.
- Capture campaign
- The campaign that collected the sequences, with its acceptance criteria.
- Dataset release
- The versioned dataset the model trained on.
- Evaluation cohort
- The held-out scenes used to score both model versions.
- Outcome
- What the evaluation measured, with confidence intervals.
- Next allocation
- Where the next round of collection goes, and why.
Diagnose
From rollout video to a collection brief.
Send your failures
Send rollout videos and logs as MCAP files, ROS bags or LeRobot datasets. Avala splits each sequence into subtasks and finds the point where it went wrong.
Read the gap report
Failures are grouped into a failure taxonomy and mapped to coverage cells. The report shows which cells are thin, which fail most often and which your data has never covered.
Approve the brief
Avala proposes a collection brief with tasks, settings, objects, sequence counts and acceptance criteria. Your team approves it before capture starts.
Every sequence
Two separate grades on every sequence.
A recording can be technically perfect and still show a failed task. Avala grades the two separately, so a sensor fault never hides a useful mistake.
Technical validity
- Sensor integrity
- Time synchronization
- Calibration
- Visibility of hands and objects
- Privacy
- Complete metadata
Behavioral outcome
- Expert success
- Partial success
- Mistake and recovery
- Intervention
- Unsafe or aborted
- Novel strategy
Recoverable failures are labeled and delivered as their own lane. They show a model how to get back on track, so Avala keeps them.
Sequence metadata
What ships with each sequence.
- Outcome label on both grades
- Subtask boundaries with a language description of each step
- Quality and speed grades
- Mistake and recovery segments
- Rig and calibration IDs
- Leakage groups by home, object and operator
Leakage groups let you split training and test sets so that the same home, object or operator never appears on both sides.
Product
Built on product you can use today.
Fleet
Devices, recordings, events, rules and alerts for every robot that reports in.
Capture campaigns
Campaigns that define what to collect and check each submission against the brief.
Quality control
Acceptance criteria, quality targets, consensus scoring and video fingerprint overlap checks.
Viewer and 4D reconstruction
The MCAP viewer and Gaussian-splat 4D reconstruction to inspect any sequence in its scene.
MCP server and SDKs
Query fleet recordings, capture campaigns, acceptance coverage and consensus from your agents and code, in Python or TypeScript.
LeRobot export
Export a dataset release in LeRobot format from the Python SDK or the command line, ready for your training code.
Where Avala sits
You own the model. Avala runs the infrastructure under it.
Avala is in the infra layer, directly under the customer’s model. We do not build robots, own the customer’s model, sell compute, or make chips. We build the data and learning infrastructure by which the customer’s own model learns from its own fleet.
Start
Bring one failure mode. Leave with a brief.
In a working session we look at your rollouts, name the failure you most want to fix, and draft the first collection brief and evaluation cohort.