Two halves of physical AI: force control decides how hard, cognition decides whether it is a defect

When a robot wipes a table, too little pressure leaves it dirty and too much makes the motion rigid; if someone bumps its elbow it still has to finish the stroke. That is no longer a vision problem. Inspection meets a threshold that looks different but is the same in kind.

ACI in this article stands for Auto Cognitive Inspection — the next-generation term for AI AOI: keep the A and the I, and replace the O (optical) with C (cognitive), because the decision basis is cognition rather than the imaging method. The full argument is set out in From AI AOI to ACI.

A domestic force-control robotics supplier recently announced that cumulative deliveries of its humanoid force-controlled arms had reached ten-thousand-unit scale, with customers reported to come mainly from embodied-intelligence teams. The figures are theirs; what matters here is the conclusion they point to — the model layer is no longer the bottleneck, getting models into the physical world is.

The proposition has two halves: execution needs a sense of proportion, cognition needs the nerve to decide.

On the execution side the question is how hard to press

Position control knows where it should go; it does not know how hard it is pressing right now. A cloth has to hold pressure while it follows a path; a carried cup needs the disturbed segment to comply while the end effector stays steady; the last few millimetres of a USB insertion are not solved by more precise coordinates. These are continuous quantities, and the answer is force.

That is why force is being written into the models: force feedback is entering multimodal training, and how closely a teleoperated arm follows the operator decides whether the motion data collected can be used for training at all.

On the cognition side the question is whether it is a defect

Industrial inspection faces a threshold of the same shape. Conventional AOI decides by comparison: match the object against a template or a rule set, flag what does not match. It knows what the thing should look like; it does not know whether what it sees now counts. That is the same gap as knowing coordinates but not force.

So every changeover, every batch-to-batch variation in incoming parts and every drift in lighting sends someone back to reprogram or retune thresholds. The software is not too slow; its decision principle has no layer for understanding.

DaoAI calls this class of system ACI: what has gone out of date in AOI is not the A or the I but the O — "optical" pins the category to one imaging modality, a visible-surface object and a compare-against-template rule.

What the last few millimetres look like in inspection

• Changeover: how long before a board never seen before can be inspected
• Incoming batches: legitimate appearance variation between batches — which of it counts as a defect
• Lighting and process drift: when the environment moves, should the decision move with it
• New defect forms: the one that never entered the sample library, the first time it appears

All four happen outside the demo. They are also where feature cognition parts company with comparison: a single good sample is enough to build a model, with no CAD drawings and no pre-collected defect library; changeover needs no code, and the model keeps iterating on line feedback. (typical achievable ranges in anonymised industry scenarios; actual figures vary with process, component and existing defect library, and on-site measurement governs)

Infrastructure is not asked to be clever, it is asked to be reliably there

The judgement from the execution side applies here too: once customers wire equipment into their own training systems, production lines and data loops, the last thing they want is one batch behaving differently. Consistency, repeatability and no drift over long runs — requirements that do not sound like selling points are exactly what decides whether a system can become infrastructure.

That is also why the inspection software is called an operating system: it has to govern imaging and motion below, present a stable decision interface above, and behave the same across models, lines and batches.

Both halves share one physical world model

Force control and cognition look like two technical routes, but they have to understand one world: space, material, light, motion and causality. DaoAI models those five in a single foundation, DaoAI World, and derives per-scenario models from it — cognition is the brain, camera and gantry motion control is the cerebellum, and the foundation keeps both consistent about the same physical world.

Robots leave the tidied demo environment and enter a world with error, friction and collision not because one side alone got stronger. The arm has to know how hard it is pressing; the eye has to know whether what it sees counts. Without either half, the model has not really entered the physical world.

Full solution for this scenario: ACI OS industry solutions

This article was generated by AI. Customer cases are simulated scenarios based on real product capabilities and figures are illustrative; see product pages for official benchmarks.