Separate intent, planning and safe motion
A multimodal model can interpret a scene, break a goal into steps and select a skill. A low-level controller should execute motion within defined limits, while an independent safety layer stops it when those limits are violated. This architecture allows the intelligence layer to change without redesigning the complete safety case for every model update.
For each skill, specify permitted objects, poses, forces, workspace, speed and failure conditions. An uncertain model output must not automatically become a physical action.
Evaluate the task, not a polished episode
Test known and novel objects, changed backgrounds, occlusion, contradictory instructions, moving people, lost visibility and failed grasps. Metrics should cover completion, attempts, time, unsafe proximity, correct stopping and the ability to recognise that a task did not succeed.
Exercise long sequences because a small per-step error quickly reduces end-to-end success. Test skill transfer between robot forms and calibrations separately.
Data and simulation need provenance
Record source, version, consent and use constraints for demonstrations, telemetry and synthetic data. Simulation can generate rare and dangerous situations, but critical behaviour still needs confirmation on physical equipment. Failure data is often more valuable than perfect repetition because it teaches detection and recovery.
Maintain traceability between data, model, skill and evaluation versions so rollback remains a practical engineering operation.
Increase autonomy in stages
Begin with recommendations to an operator, then execute a step after approval, move to a constrained zone and grant wider autonomy only after evidence accumulates. Escalation thresholds should reflect the consequence of an error, not average model accuracy.
Production readiness means the system knows its skill boundaries, fails safely, preserves a complete decision trace and can return to a previous version without prolonged process downtime.