Task Adaptation
New task. No reprogramming.
The floor supervisor performs the target task once — or up to three times for tasks with significant positional variation. Atom records joint trajectories, applied force profiles, and sensor readings throughout. That recording is processed into a motion policy: a compact generalised representation of the task that tolerates normal variation in object placement. Deploy it the same shift. No programming, no systems integrator, no change order.
The Process
How demonstration-based learning works in practice
01
Supervisor demonstrates
The floor supervisor performs the target task — physically moving objects as they would during normal operations. Atom records joint trajectories, force profiles, and sensor readings throughout.
02
Policy embedding
The recorded demonstration is processed into a motion policy — a compact representation of the task that generalises across minor positional variation in object placement. Typically requires 1-3 demonstration runs for robust policy generation.
03
Supervised deployment
The supervisor reviews the policy before deployment on a verification run. If the policy fails the verification criteria, the system flags it for re-demonstration rather than auto-deploying a low-confidence policy.
What it handles well
Routine task variation
Demonstration-based learning handles tasks where the goal is stable but the exact object placement varies within a range the policy can generalise across. Assembly subcomponent transfer, bin-to-conveyor routing, kit preparation, pallet sorting by destination label.
Multi-step tasks work: the supervisor demonstrates the complete sequence, and the learned policy includes the dependency ordering between steps — Atom does not need to rediscover it at execution time.
Current limits
Where human operators still lead
Tasks that require force application finer than the current torque resolution of 0.1 Nm — precision fastener seating to sub-degree accuracy, for example — are at the edge of the current capability envelope. High-payload tasks above 10 kg continuous are outside it. We describe these limits clearly at the briefing stage rather than discovering them during a trial deployment.
Atom is not a quality inspection system. It can position a component for a human inspector or an optical measurement device — it cannot substitute for a trained quality inspector making a conformance judgement.