Ingest & Visualize: inspect your sensor recordings
Prepare a supported recording, upload it as a private dataset, and check timing, calibration, and coverage before labeling.
A recording may contain cameras, LiDAR, an IMU, and robot state at different sample rates. Before labeling it, confirm which channels are present, how timestamps are represented, and which transforms connect the sensors.
Prepare a supported recording
MCAP packages messages and their schemas in a container suited to sensor recordings. If your data starts as a ROS bag or another format, prepare a supported MCAP recording using the appropriate conversion tooling before uploading it. Do not assume that changing the file extension converts the data.
The CLI supports local uploads for image, video, LiDAR, MCAP, and splat datasets. The following command uploads your file to Avala-managed storage and creates a new private dataset. Replace the source path with a real recording and choose a dataset slug that is not already in use. It does not append data to an existing dataset.
avala datasets upload \
--source ./recordings/run_42.mcap \
--name "Fleet recording" \
--slug fleet-recording \
--data-type mcap \
--visibility private \
--wait
Add --dry-run to inspect the local file selection before uploading. --wait polls for dataset indexing after the upload. If you are using a cloud storage integration, follow that integration's setup rather than assuming this local-upload command connects a bucket.
Inspect timing and geometry
Open the recording in the sensor viewer and inspect the channels your data provides. Use the timeline to compare events across available camera, point-cloud, and state views. WebGPU supports browser-based rendering, while the usable views depend on the recording's message types and configuration.
Check the data rather than assuming it is synchronized:
- Timing: identify offsets, dropped frames, and timestamp conventions.
- Transforms: verify the coordinate frames and sensor-to-sensor transforms.
- Calibration: compare projected geometry with camera images at several moments.
- Coverage: confirm that the recording includes the event and context you need to label.
Gaussian splats are a separate representation with their own reconstruction inputs and processing. Uploading an MCAP file does not guarantee that a splat reconstruction will be generated.
Decide what is ready to label
A useful inspection ends with a decision: which sequences are ready, which need repaired metadata or calibration, and which moments belong in the annotation scope. Keep the source recording and the relevant context available to reviewers.
The viewer helps you find these problems. It does not make inconsistent source data correct on its own.