ou need to annotate 20,000 frames of robot video over 9 months. Two approaches compete:
Budget-conscious approach: Hire gig workers on a task-by-task basis. Turnaround is fast, headline cost is low. By month 4, you're reannotating 30% of frames because workers misunderstood the rubric. By month 6, you're updating the schema and retraining workers. By month 9, invisible rework costs have accumulated invisibly.
Specialist approach: Hire 4 dedicated specialists for the full 9 months. They annotate 10,000 frames and know your robot, your annotation rubric, and your edge cases by heart. They're producing 98% accuracy on first pass. No rework cycle.
This is the retention economics of multi-quarter annotation programmes.
The ramp-up curve: where turnover kills you
New annotators are unproductive for 4–6 weeks:
Week 1–2: Learning the tool, understanding the rubric Week 3–4: Reaching 70% accuracy Week 5–6: Reaching 90% accuracy Week 7+: Reaching 95%+ accuracy, developing intuition about edge cases
If you hire fresh workers every 4 weeks (high turnover), you have no one past week 6. Your team average productivity is stuck at 70–80% accuracy and 50% throughput.
If you hire 4 specialists and keep them for 9 months, weeks 1–6 happen once. From week 7 onward, you have 4 people at 95%+ accuracy producing work at 3–4x the rate of novices.
The math:
- 4 novice workers, 4-week churn cycle: ~800 frames/month at 75% accuracy
- 4 seasoned specialists, 9-month engagement: ~3200 frames/month at 95% accuracy
Over 9 months: Novice path produces data that requires extensive rework cycles, invisible costs mounting throughout the programme. Specialist path produces data that is usable on first pass, eliminating rework overhead and compounds expertise through iteration.
Building and retaining annotation teams
Hiring: Look for people with 2+ years robotics industry experience, CAD software proficiency, or engineering background. They learn your rubric 2–3x faster than generalists.
Onboarding: Invest 2–3 weeks in structured training. Create a rubric walkthrough, show them common errors annotators make, have them practice on 100 frames before production work. This investment pays back within 2 months.
Compensation: Offer competitive rates that signal stable, valuable work. This improves retention and quality, which reduces invisible rework costs.
Stability: Offer 6–12 month contracts with 2–4 week notice periods. This signals to annotators that they have stable work, which improves retention.
Skill development: As the programme evolves, promote annotators into QA roles. Your best annotator becomes your QA lead, training new hires and reviewing hard frames. This improves retention (career path) and reduces rework.
Iteration and schema evolution
Multi-quarter programmes inevitably need schema changes. Your initial rubric covers 80% of cases; months 3–4 reveal edge cases that require rubric refinement.
With gig workers, this is a disaster: you announce a rubric change, workers reinterpret it differently, and accuracy tanks.
With a stable team, rubric evolution is manageable: you meet with your 4 specialists, discuss the edge case, agree on an interpretation, and roll out the change. They understand the intent, not just the letter.
Common schema changes in months 2–6:
- "How do we label occluded gripper state?" (you didn't think of this initially)
- "Should we label the shadow of the robot arm, or just the arm?" (ambiguous edge case)
- "Contact points: do we label them when the gripper is closed, or when it first touches?" (temporal definition)
Each of these requires one discussion with your core team, not retraining 50 workers.
Quality and velocity tradeoff
Gig platforms optimise for velocity: "We'll annotate your 10,000 frames in 2 weeks." The catch: at 70% accuracy, 3,000 frames need rework.
Specialist teams optimise for quality: "We'll annotate 1,000 frames per week at 95% accuracy." This seems slow until you account for zero rework.
Real timeline:
- Gig path: 2 weeks annotation + 6 weeks rework + 2 weeks retraining = 10 weeks
- Specialist path: 10 weeks annotation + 1 week final QA = 11 weeks
Same timeline, but specialist path produces usable data. Gig path produces data that requires full re-annotation.
Cost of frequent platform changes
If you swap platforms every 2 months, you lose expertise in each tool's workflows. Annotators re-learn hotkeys, UI navigation, and export formats. This adds 15–20% overhead per platform switch.
Stable teams stay on one platform for the full 9 months, optimising their workflow, building macros, and developing muscle memory.
What this means for you
For any annotation programme >3 months, retention beats price. Hire a small team of specialists, commit to 6–12 month engagements, invest in onboarding, and iterate your schema with their input.
This is especially critical for robotics, where domain expertise directly impacts label quality. You're not buying labour; you're buying encoded knowledge about your system.
The companies that deploy models fastest aren't the ones with the cheapest annotation — they're the ones with stable, experienced annotation teams. Every day a specialist doesn't have to re-learn your rubric is a day they spend improving data quality. Better data means faster training cycles and more reliable models in the field.
Plan for retention as a core part of your programme. It's the single biggest lever on cost, timeline, and model quality.
Explore robotics data programmes or discuss long-term partnerships with our team.