We evaluate our method on five challenging tool-use tasks specifically designed to isolate finger dexterity from arm motion. The robotic arm remains stationary, while all manipulations are executed solely by finger movements. Each task is repeated ten times, and success rates (SR) are reported. As summarized in the table above, our approach demonstrates strong improvements after fine-tuning, achieving an overall success rate of 84.0%, compared to only 36.0% before fine-tuning.
The tasks cover diverse tool manipulations, including chopsticks, brush, syringe, scissors, and spray bottle. Before fine-tuning, performance was limited because the operator had to adopt unnatural or counterintuitive finger gestures, resulting in unstable grasps and frequent failures. For instance, chopsticks and scissors tasks had only a 10% success rate due to the difficulty of consistently reproducing morphology-adapted grasps.
After fine-tuning, our networks reinterpret natural, operator-friendly finger gestures into robot-specific motions adapted to the hand morphology. As a result, success rates increased substantially across all tasks: 60% for chopsticks, 90% for brush, 100% for syringe, 90% for scissors, and 80% for spray bottle use. These results confirm that fine-tuning effectively compensates for human–robot morphological differences, reduces operator burden, and enables robust execution of challenging tool-use tasks.
