A person can pick up an unfamiliar cup while talking, adjust grip when it slips and place it on a crowded surface. The action feels ordinary because perception and movement have been refined through years of embodied experience. For a robot, the same task joins several difficult problems: recognising the object, estimating its shape and weight, planning a collision-free motion and adapting when reality differs from the model.
Sensors provide partial views
Cameras flatten a three-dimensional scene and change with lighting. Depth sensors can be confused by reflective or transparent materials. Touch arrives only after contact. No single sensor delivers a complete description of the environment.
A robotic system must combine uncertain signals while deciding which information matters for the next action. It also needs to recognise when confidence is too low. Mistaking uncertainty for absence is dangerous; an object that was not detected has not necessarily disappeared.
Action changes the evidence
In a static image task, analysis does not alter the picture. A robot changes its environment whenever it moves. Reaching toward an object may push another one, obscure the camera or place the machine in a configuration with fewer safe options.
Planning must therefore consider sequences and recovery. The best immediate movement may be inferior if it leaves no graceful response to a slip. Robust control continuously compares expected and observed motion, making small corrections rather than assuming the original plan remains valid.
Everyday environments contain long tails
Homes, streets and workplaces are full of variations that are individually uncommon but collectively normal: loose fabric, uneven ground, unusual packaging, a person stepping backwards. Collecting examples helps, but designers cannot list every possible exception.
Constrained deployments are often the responsible path. A robot can deliver value in a well-defined environment with known objects and clear separation before expanding to less structured settings. Capability should grow alongside evidence about failure, not ahead of it.
Physical intelligence is hard because the world does not wait for a perfect model. Useful robots need perception that admits uncertainty, control that adapts during action and limits that keep unfamiliar conditions from becoming unsafe experiments.