Define the task in observable terms
Start with a clear initial state, permitted actions and a completion condition. For a checkout item, that might include a successful scan and correct placement. For replenishment, it includes the correct shelf and item orientation. Define the conditions where the robot must stop or request assistance.
FastLane structures training around those retail requirements. Demonstrations, simulation and learned control are tools in that process. Which tools are appropriate depends on the robot’s control interfaces, available data and the manipulation problem.
From demonstration to supervised trial
- Capture the workflow. Document staff actions, product variation, interface states and recovery procedures. Decide which observations are actually needed.
- Collect examples. Where the platform supports it, teleoperation can supply demonstrations. Good examples include corrective actions and clear task boundaries.
- Use simulation deliberately. Model the relevant geometry and vary lighting, positions, object properties and timing where those parameters matter.
- Test unseen conditions. Hold out products, arrangements or recording sessions so the evaluation is not simply a replay of the training set.
- Validate on hardware. Check calibration, contact behavior, perception and stopping on the actual robot and workstation.
- Run the bounded pilot. Measure task completion, intervention and recovery with the agreed supervision, then decide what to improve.
Simulation accelerates learning; hardware settles the contact problem
Virtual environments are useful for repeated controlled variations. Grocery packaging adds difficult physical effects: deformable pouches, reflections, transparent trays, loose contents and changing friction. Real equipment also brings wear, latency and calibration error. Those effects must be tested rather than assumed away.
The sim-to-real guide explains how imitation learning, reinforcement learning and domain randomization fit together, with a practical evaluation matrix.
Teleoperation has more than one role
Direct remote control can collect demonstrations. Remote assistance can resolve a bounded exception. Supervisory approval can authorize a discrete action. These modes require different control interfaces, staff skills and response-time assumptions.
For an operational project, define who may take control, what they can see, how authority transfers, what happens when the connection fails and when a person on site must intervene. Remote support does not replace the local protective measures and risk assessment for the workcell.
Ask for evidence that survives a longer trial
- Total attempts and completed tasks, with the tested conditions.
- First-attempt success separated from success after recovery.
- Human minutes spent supervising, intervening and resetting.
- Cycle-time distribution rather than only the fastest run.
- Failure categories, damaged items and clear stop events.
- The data and software rights needed to maintain the workflow.
Continue with teleoperation in retail or the pilot evaluation guide.