Computer vision has made robots far better at locating objects and interpreting scenes, but an image does not reveal everything that matters during physical manipulation. A camera may estimate the pose of a component while remaining unable to confirm whether the part slipped, resisted insertion or was gripped too firmly. Robots operating outside tightly controlled fixtures therefore need information from contact as well as sight.
Vision ends where physical interaction begins
Cameras can identify parts, estimate position and guide motion toward a target, yet their performance depends on lighting, occlusion, calibration and how closely the scene matches expected conditions. Even an accurate pose estimate cannot confirm mechanical success. A connector may look aligned and still bind during insertion because of a small tolerance stack. The robot needs feedback that describes what happens after contact rather than assuming that visual alignment guarantees the outcome.
Contact signals give robots a second kind of evidence
Force and torque measurements can reveal first contact, resistance, slip, off-centre loading or whether a part has reached an expected mechanical condition. Gripper width, motor current and tool-state signals add further clues. The useful learning signal is not merely what the robot sees, but how the physical world responds to the action it takes. That closes the loop between perception, movement and measurable task outcome.
Control has to react during the task
A precomputed path assumes that the world stays as modelled. Force-aware control can instead adjust motion when contact differs from expectation, for example by reducing speed, changing direction or stopping when resistance exceeds a defined threshold. This can help with insertion, finishing and other contact-rich operations, but it is not equivalent to human touch. Thresholds and controller behaviour must be tuned to the mechanics of the tool, object and robot or the system can oscillate, damage parts or stop unnecessarily.
Learning requires outcomes, not just sensor streams
A force trace shows that something happened, but training is more useful when the system also knows whether the grasp remained stable or the insertion passed inspection. Machine signals, tool states and quality checks can provide these labels. Synchronisation matters because camera frames, joint positions and force readings must share a reliable timeline if a model is expected to connect an action with its consequence. Poor calibration can teach the wrong relationship even when every sensor appears to work on its own.
The tool defines what can be sensed and learned
A wrist sensor measures overall loads, while fingertip sensing may reveal local contact and a vacuum tool exposes a different set of states again. For physical AI this matters because the model can only learn from physical variables that the cell is actually able to observe, and each end effector also constrains the actions available to the robot. A two-finger gripper, suction cup and sander do not create the same learning problem even when mounted on the same arm, so sensing and action design must be treated together.
Trust grows from observable behaviour
Industrial learning systems still operate inside machines with momentum, pinch points and process hazards. Their permitted actions therefore need independently enforced safety limits, and new learned behaviour must be validated before it controls production motion. Vision remains important because it provides context before contact, while force and tool feedback provide evidence during interaction. Robots become easier to trust when they can show through measurable signals not only what they intended to do, but what their actions actually achieved.
Physical data also changes maintenance
Interaction data can support engineering even when it is not used to train a model. A gradual rise in grip current, contact force or failed insertion attempts may indicate tool wear, fixture drift or a change in incoming parts before a complete stoppage occurs. When those signals are linked to maintenance records and quality results, the cell becomes easier to diagnose and improve. Physical intelligence can therefore deliver value through observability as well as through autonomous adaptation.
