Putting a small network connector into its socket is harder than it looks. The plug must face the right way, line up and enter without slipping. A team led by Ammar Abbas compared ways to teach a five-fingered robot hand this task. They carried out roughly 1,700 attempts on real equipment, not just on a screen.

The team tested 281 versions of control software from seven groups. They changed what information the program received from cameras and how it described finger movements. This is called a data representation: the same task can be described using different pictures and numbers. Sometimes that choice mattered more than ordinary adjustments to the training process.

The researchers also examined 67 scores that could be calculated without performing the insertion. Only five, related to how smooth and direct the predicted motion was, closely matched real success. Even those worked only within one group of programs. A movement that looks excellent in a computer test may still miss the actual socket.

For point-cloud^[1] policies, cropping the relevant region, using robot-base coordinates and normalizing features improved success by 70 to 100 percentage points in individual controlled comparisons. For a vision-language-action policy, absolute rather than relative actions improved insertion success by up to 70 points. These figures come from different configurations and cannot be added together or generalized to every architecture. Seven policy families were included, exposing how representation choices interact with model design.

The findings concern particular hands, connectors and test conditions. They do not establish the best program for every object. They do explain why a robot must be tested with the things it will really handle.

Only five of 67 offline measures reliably correlated with real insertion success, and only within one policy family. Correlations between minus 0.70 and minus 0.80 for predicted motion smoothness and directness do not turn these scores into substitutes for hardware trials. Selecting a controller from validation loss alone is particularly weakly supported here.