Physical AI earns its place through a complete operating loop

A successful pick is a demonstration. A production system must also recognize a failed pick, leave the process in an appropriate state and help an operator recover.

Specify the whole task

Take an illustrative inspection-and-sorting cell. Parts arrive with variable position and appearance. The cell must identify the part, estimate its state, select an action, execute the move, inspect the result and record the outcome. The requirement includes the exceptions: no part, damaged part, uncertain inspection, blocked destination and communication loss.

Define completion in terms the customer can accept. Record cycle-time distribution, accepted and rejected results, interventions, recovery time and downtime. A test set should include expected variation and deliberately introduced disturbances, with the procedure and limits agreed in advance.

Keep responsibilities explicit

A task planner can select a permitted sequence; motion planning produces a feasible path; real-time control follows the motion command. Safety functions and authorised operating states need their own engineering basis. A language model’s confidence is not a safety mechanism.

Dyna’s published work describes a particular integration of learned behavior and whole-body control. UMI research shows how task demonstrations can be captured with a compatible gripper interface. Both are useful technical references; neither eliminates application-specific integration and validation. Dyna 2.1 ↗ · Universal Manipulation Interface ↗

Collect evidence that explains failures

An episode should connect observations, actions, task state, configuration and result. Keep hardware revision, calibration, software and model versions with it. Synchronize timestamps and record missing data so a replay does not suggest certainty the instruments did not provide.

Hyphenbox describes source-linked annotations for subtasks and failure spans. The useful principle is traceability: retain the original evidence and distinguish automated annotation from human review. An attractive timeline does not make its labels correct. Hyphenbox annotations ↗

Design for the technician after handover

Provide local event buffering, recoverable configuration, clear fault explanations and approved operator actions. Decide what continues during a network outage. Protect access to the equipment and document who can change a controller, model or safety-relevant setting.

The first deployment can implement these basics without a generalized fleet platform. When a field failure becomes a repeatable regression test, the next installation inherits something valuable. SSA’s Deploy programme is being developed around this task-level responsibility; the interactive workcell is a conceptual exhibit, not a report of installed performance.