Leadership and readiness
Robotics data collection downtime costs: a worksheet for host sites
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Estimate the cost of hosting workplace video collection in hours before you price anything: station time lost to installation and slower work, staff time for briefings, escorts and demonstrations, per-episode overhead from resets and retakes, and review time. Give each input a low, base and high value. Go ahead only if the high case still fits your production plan.
Video and sensors
Robotics data collection
Key takeaways
- Most of the cost of hosting collection is time: stopped or slowed stations and staff hours, not equipment.
- Mounting cameras on or near machines can count as servicing work, so plan it inside a maintenance window.
- Reset time between recorded attempts can match the recording time itself, and retakes add to both.
- The throughput loss while cameras run is often the input that changes the decision; measure it in a short pilot.
- Footage of only clean runs is cheaper to record and worth less, because buyers look for variation and recovery.
How do you estimate the operating cost of workplace video collection?
Count it in hours first and convert to money last, using your own rates. A host site pays in two currencies: station time, meaning the hours a line or cell produces less or nothing, and staff time, meaning the hours people spend on the program instead of their normal work. Both are known to your operations team; neither depends on what a buyer or collector believes.
The decision rule is simple. Build the estimate in five buckets, give each input a low, base and high value, and go ahead only if the high case fits inside the downtime your production plan can absorb and the staff hours your supervisors agree to release. If only the base case fits, run a short pilot to measure the uncertain inputs before committing to the full program.
Where the cost of hosting video collection comes from
The cost sits in five buckets, and only one of them is visible on a collector's proposal. Setup and installation is the obvious one. The others are lost output while cameras run, people time for briefings and supervision, the overhead of recording attempts, and the review and preparation work that follows.
At a glance
Where the cost of hosting video collection comes from
Collection is labor-heavy even for experienced teams. The DROID robot manipulation dataset, for example, reports roughly 76,000 demonstrations and 350 hours of interaction data collected by 50 data collectors across 564 scenes over 12 months. A host site is only one scene in such a program, but the hours per usable episode are just as real.
- Setup: site survey, mounting and calibrating cameras, network and storage, removal afterward.
- Production impact: stations stopped during installation and slowed while workers are recorded.
- People: worker briefings and consent conversations, supervisor and safety escort time, demonstrator time.
- Episode overhead: reset between attempts, cancelled and re-recorded attempts, human interventions.
- Review and preparation: the company's own review of what was filmed and the task-boundary labels it signs off.
The downtime and preparation cost worksheet
Fill one row per input with your own values. The illustrative values are the ones used in the worked scenario below, not benchmarks. Formulas give station-hours (SH) for production impact and staff-hours for people time, kept separate because they are priced differently.
Two rows are easy to undercount. Briefings happen again for every shift and every new hire who joins during the program, not once. Review sign-off grows with camera count, because each added angle multiplies the footage someone at the company has to sample before release.
| Input | Definition | Formula | Illustrative base value |
|---|---|---|---|
| Install downtime | Hours each station is stopped for mounting, cabling and calibration | stations x stop hours per station = SH | 2 stations x 1.5 hours = 3 SH |
| Recording slowdown | Share of normal output lost while recording, from pacing, retakes and attention | stations x recording hours x slowdown % = SH | 2 x 21 hours x 10% = 4.2 SH |
| Maintenance and IT time | Technician hours to mount, test and later remove equipment | sum of task hours | 4 staff-hours |
| Briefings and consent | Time each affected worker spends in notice, questions and consent | workers x (briefing + consent hours) | 12 x (0.5 + 0.25) = 9 staff-hours |
| Supervision and escort | Staff time accompanying the collection crew and keeping safety rules | days x escort hours x share of attention | 3 x 8 x 50% = 12 staff-hours |
| Demonstrator time | Attempts times (episode length + reset), with retakes | target episodes x (1 + retake %) x (episode + reset) | 300 x 1.15 x 2 minutes = 11.5 staff-hours |
| Supervisor planning | Scheduling, task list approval and daily check-ins | sum of task hours | 4 staff-hours |
| Company review sign-off | Sampling the footage for screens, badges and out-of-scope areas before release | footage hours x sample % x review ratio | 17.25 x 10% x 1.5 = about 2.6 staff-hours |
| Contingency | Allowance for incidents, equipment faults and repeat days | subtotal x contingency % | 15% on both SH and staff-hours |
Mounting cameras can stop the line: plan it as maintenance
Treat any mounting on or inside a machine guard as servicing work and schedule it in a planned maintenance window. OSHA's lockout/tagout standard, 29 CFR 1910.147, applies to servicing and maintenance where unexpected startup or release of stored energy could injure employees, and its definition of servicing includes installing, setting up and adjusting equipment.
That has a direct cost effect: if mounting needs the machine locked out, the station is down for the whole job, not just the minutes of drilling. Tripod and overhead mounts that stay outside guarded areas often avoid the stop entirely, at the price of a worse viewing angle. Ask your safety lead which mounts on which machines need lockout before the collector designs the camera plan, because the answer can move install downtime from near zero to several station-hours.
Per-episode overhead: resets, retakes and interventions
Budget demonstrator time per attempt, not per minute of usable footage. Open-source collection tooling makes the overhead explicit. LeRobot's recording command defaults to 60 seconds per episode, 60 seconds of reset time between episodes and 50 episodes per session, with keyboard controls to cancel an attempt and re-record it. With those defaults, reset time equals recording time.
Retakes stack on top. If one attempt in seven is cancelled, every usable episode costs more than its own length plus reset. Human-in-the-loop collection adds another category: LeRobot's documentation describes recording human interventions, the recovery movements and corrections, alongside autonomous segments. Those segments are valuable, but they need a trained person standing by for the whole session.
For a host site, the practical question is who the demonstrator is. If it is a production worker, demonstrator hours are also lost production hours; if it is the collector's own operator, they become escort hours for your staff instead.
Worked scenario (illustrative): two packing stations over three days
Everything in this scenario is invented to show the arithmetic. A packaging company hosts collection at two packing stations for three days, with seven recording hours per station per day, two fixed cameras and one head-mounted camera, 12 affected workers and a target of 300 usable episodes recorded by one trained demonstrator from the company's own staff.
Station time: 3 SH of install downtime plus 4.2 SH of slowdown gives 7.2 SH, or about 8.3 SH with 15% contingency. Staff time: 4 maintenance, 9 briefing and consent, 12 escort, 11.5 demonstrator, 4 planning and 2.6 review hours give 43.1 staff-hours, or about 49.6 with contingency.
Finance then prices the two totals separately: station-hours at the margin the stations would have earned, staff-hours at the loaded hourly cost of the people involved. Use loaded cost, not the wage rate: the Bureau of Labor Statistics measures employer compensation cost per hour worked as wages plus benefits, and in June 2026 wages and salaries were 70.0 percent of private-industry employer costs. The company then compares the result with what the agreement covers and with the disruption its customers would notice, and decides.
Which assumption changes the decision, and how to show the uncertainty?
Run each uncertain input at a low and high value while holding the others at base, and look for the one that moves the total most. In the scenario, the recording slowdown dominates station time and the reset ratio dominates demonstrator time. Present the result as a range, not a point, and say which input a pilot day would pin down.
A pilot day is the cheapest way to narrow the range. Record one station for a full shift with the planned camera set, count output against the same shift on a normal day, time every reset and log every cancelled attempt. Those three measurements replace the three most uncertain inputs with observed values from your own floor.
Watch for inputs that move together. A cramped station often raises both the slowdown and the reset time, and a new task list raises both retakes and escort attention. When two inputs share a cause, run them high together as a stress case instead of trusting the one-at-a-time table alone.
| Input varied | Low value and result | Base value and result | High value and result |
|---|---|---|---|
| Recording slowdown | 5%: 2.1 SH | 10%: 4.2 SH | 25%: 10.5 SH |
| Reset time per attempt | 30 seconds: about 8.6 demonstrator hours | 60 seconds: 11.5 demonstrator hours | 120 seconds: about 17.3 demonstrator hours |
| Retake rate | 5%: 10.5 demonstrator hours | 15%: 11.5 demonstrator hours | 40%: 14 demonstrator hours |
| Install stop per station | 0 hours with external mounts: 0 SH | 1.5 hours: 3 SH | 4 hours with lockout and rework: 8 SH |
| Escort attention share | 25%: 6 staff-hours | 50%: 12 staff-hours | 100%: 24 staff-hours |
Footage without failures and recoveries costs more later
Plan for variation, failures and recoveries in the task list, even though they make each day slower. The cheapest footage to record is a skilled worker doing the clean case repeatedly, and it is the least useful, because robots fail on the jams, mislabels and awkward items the clean case leaves out. A second collection visit to capture those cases costs more than recording them the first time.
Write the variation into the episode plan: a share of episodes with damaged packaging, a share with a missing item, a share where the worker recovers from a dropped object. Record the outcome of each attempt and any intervention, because footage with outcomes can be used for evaluation as well as training. Large research collections such as Ego4D describe collection from consenting participants with de-identification procedures; the same discipline applies when the people on camera are your own staff.
Worker notice, consent and labor rules belong in the estimate
Budget the hours for notice and consent as a fixed line, and have counsel review the program before cameras arrive. Rules on workplace recording, biometrics and audio vary by state and by country, and they affect who may be filmed, what must be blurred and whether audio may be captured.
Labor law also matters in the United States. An NLRB General Counsel memo on electronic monitoring, GC 23-02, named cameras and wearable devices among the technologies of concern; it was rescinded in February 2025 by GC 25-05, but the National Labor Relations Act itself did not change. Workers who decline to take part need a real alternative, which can mean scheduling around them and adds hours to the estimate.
Licensing footage you already have versus hosting new collection
Before hosting anything, check whether existing records already answer the buyer's need. Footage from safety, training or quality cameras, linked to work orders and incident records, can sometimes be licensed with review and preparation but no new downtime at all, though recordings made for other purposes need their own rights and privacy review.
Licensing existing footage is the kind of project SourceX runs: it takes a company's selected records through rights review, contracting and preparation until they become a licensed AI dataset. The first step is a short company assessment, with no data uploads needed to see whether a company qualifies.
Frequently asked questions
What is the biggest hidden cost of hosting robot data collection?
Often the slower pace of work while cameras run, because it applies to every recorded hour. Installation downtime is larger per event but happens once. Measure the slowdown in a pilot day before relying on any estimate.
Do camera installations require lockout/tagout?
They can. OSHA's lockout/tagout standard covers servicing and maintenance, which includes installing and setting up equipment, where unexpected startup could injure someone. Your safety lead should decide per machine and mounting method.
How much reset time should a collection plan assume?
Assume reset time is comparable to the recording time until a pilot shows otherwise. Open-source tooling such as LeRobot defaults to equal episode and reset times, and retakes add further attempts on top.
Should the worksheet include money amounts?
Keep the worksheet in hours and lost station time, then let finance convert the totals with the company's own loaded labor costs and margins. That keeps the operating estimate checkable by the people who run the floor.
Can we avoid downtime by licensing existing footage instead?
Sometimes. Existing safety, training or quality footage linked to work records can carry similar value with no new recording, but it still needs rights, privacy and buyer-fit review before any license.
Sources
- OSHA's lockout/tagout standard covers servicing and maintenance, including installing, setting up and adjusting machines, where unexpected startup could injure employees. Source
- LeRobot's recording command defaults to 60 second episodes, 60 second resets and 50 episodes, with controls to cancel and re-record an attempt. Source
- LeRobot human-in-the-loop collection records human interventions, recovery movements and corrections, alongside autonomous segments. Source
- The DROID dataset reports about 76,000 demonstrations and 350 hours of data collected by 50 data collectors across 564 scenes over 12 months. Source
- Ego4D describes collection from consenting participants with de-identification procedures where relevant. Source
- The BLS Employer Costs for Employee Compensation release measures employer cost per hour worked as wages plus benefits; wages and salaries were 70.0 percent of private-industry costs in June 2026. Source
- The NLRB General Counsel's memo GC 23-02 on electronic monitoring named cameras and wearable devices among monitoring technologies. Source
- GC 25-05, issued in February 2025, rescinded GC 23-02 on electronic monitoring along with other memos. Source
See if your company qualifies
A short company assessment. No data uploads are needed.