On this page
- Why most collection campaigns miss what matters
- Start from failure modes and coverage gaps
- Turn gaps into coverage targets
- Routes, schedules and conditions
- Pre-drive checks prevent wasted hours
- Grade every log at ingestion
- Report captured hours against plan
- Budgeting hours, rigs and storage
- Who should run campaign operations
- ✓Start from the model’s failure modes and coverage gaps, not from the routes that are easiest to drive.
- ✓Turn gaps into measurable targets by scenario, condition, geography and time of day.
- ✓Run pre-drive checks and grade logs at ingestion so bad hours never reach labeling.
- ✓Report captured hours against targets every week and move budget to the gaps that remain.
The short answer: A good driving data collection campaign starts with the scenarios the model handles badly, turns them into coverage targets by road type, weather, lighting, traffic and geography, plans routes and schedules to hit those targets, checks every rig before it leaves, grades every log at ingestion and reports captured hours against plan weekly. Teams that outsource the operations side usually do so through sensor and LiDAR data collection outsourcing, keeping drivers and the sensor stack in-house.
Why most collection campaigns miss what matters
Collection fleets tend to drive where it is convenient: near the depot, in daylight, on familiar roads, during working hours. The result is a dataset heavy with easy driving and light on the moments that break models, such as night rain, unprotected left turns, construction zones, emergency vehicles and unusual road users. Every extra hour of easy highway driving costs storage, transfer and possibly labeling, without improving the model much.
The fix is to plan collection the way a product team plans a release: start from the problem, define measurable targets, track progress and adjust. That turns collection from a fleet activity into a data strategy.
Start from failure modes and coverage gaps
The best input to a campaign plan is the model’s own performance. Disengagement reports, evaluation failures, safety driver notes and simulation results show which situations cause trouble. Compare those against the existing dataset to find gaps: scenarios that are rare in the data, cities or road types never driven, weather and lighting conditions under-represented, and sensor configurations with little coverage.
Write the gaps down as a ranked list with the reason each matters. A gap that causes safety-relevant failures ranks above a gap that only affects comfort. This list becomes the brief for the campaign, and it should be shared with everyone who plans routes and schedules drivers.
Turn gaps into coverage targets
Gaps become useful when they are measurable. Each target should name a scenario or condition, a geography, a time window and an amount, such as hours, events or segments. For example: forty hours of urban driving in heavy rain at night across three cities, or two hundred unprotected left turns with oncoming traffic. Rare events may need special routes, private test areas or staged scenarios.
Targets also need a definition of done. Decide in advance how a segment will be tagged as meeting the target, so progress can be counted automatically from ingestion tags rather than estimated from drive logs.
Routes, schedules and conditions
Campaign coordinators turn targets into routes and schedules. Weather targets mean watching forecasts and scheduling drives into the conditions, not around them. Lighting targets mean night and dawn shifts. Geographic targets may mean shipping rigs to new cities, arranging permits and briefing local drivers. Traffic targets mean driving at rush hour in dense areas.
Each drive should have a stated purpose tied to targets, so drivers and coordinators know what success looks like. Drives without a purpose drift back to convenient routes. See autonomous vehicles outsourcing for broader program operations around fleets.
Pre-drive checks prevent wasted hours
Many wasted hours start before the vehicle leaves: a dirty lens, a loose connector, nearly full storage, a clock out of sync or a failed firmware update. A short pre-drive checklist, run by the driver and confirmed remotely from rig telemetry, catches most of these. Rigs that fail the checklist stay in until the issue is fixed.
Track checklist failures by rig over time. A rig that fails often, or produces many rejected logs later, needs maintenance attention. That record helps hardware teams prioritize and prevents the same rig wasting a week of drives.
Grade every log at ingestion
When logs arrive, check them before anyone uses them. Confirm every expected file is present, durations match the drive plan, frame rates are stable, no sensors dropped out and calibration and time sync are within tolerance. Grade each segment as usable, partially usable or rejected, with a reason. Then tag usable segments with scenario and condition so they count toward targets.
Grading at ingestion means no team pays annotators to discover that a camera was offline. Labeling quality checks are a later stage; see our guide to LiDAR annotation quality checks for what happens once graded data reaches annotation.
Report captured hours against plan
A weekly report should show, for each target, the hours or events captured, the share rejected and why, and what remains. It should also show rig health and the drives that produced little usable data. Leaders use it to move budget: more night shifts, a new city, a staged scenario day or fewer highway runs.
Over a campaign, the report becomes the record of what the dataset contains and why. That record is valuable later, when engineers need to explain model behavior or plan the next campaign. See sensor and HD map annotation for how curated data moves into labeling.
Budgeting hours, rigs and storage
A campaign budget is more than vehicle hours. It includes driver and safety operator time, rig maintenance, fuel or charging, permits for test areas, storage for raw logs, transfer costs and the downstream cost of grading, tagging and labeling. Rare-scenario targets often cost far more per useful hour than common ones, because many drives return little qualifying data. Planning should estimate yield per target, meaning how much usable, qualifying data each drive is likely to return, and set budgets accordingly.
Reviewing yield weekly lets leaders stop spending on targets that are already full and shift money to the ones that are lagging. A target with very low yield may be better served by staged scenarios on a closed course or by simulation than by more public-road driving.
Who should run campaign operations
Campaign operations combine planning, monitoring and data handling, and they rarely fit inside an engineering team’s day. Many autonomy companies keep drivers, rigs and the sensor stack in-house, and give planning support, ingestion grading, calibration checks and scenario tagging to a dedicated operations team that works through the overnight hours when logs arrive.
Our sensor and LiDAR data collection outsourcing page explains how vetted Philippine teams run these operations, and our guide to remote operations centers for autonomous systems covers the wider operations model in the AI, robotics and AVs cluster.
Frequently asked questions
How should a driving data collection campaign be planned?
Start from model failure modes, turn gaps into measurable coverage targets, plan routes and schedules to hit them, check rigs before each drive and grade logs at ingestion.
What is a coverage target?
A measurable goal naming a scenario or condition, geography, time window and amount, such as hours of night rain driving in specific cities.
Why grade logs before labeling?
Because labeling broken or incomplete logs wastes money; grading at ingestion removes bad segments early.
How often should coverage be reported?
Weekly, showing captured hours per target, rejections by reason and remaining gaps.
