Robots may soon tell a tap from a push without any added hardware.
Researchers tested several convolutional neural network architectures on a Franka Emika Research robot, using only the joint sensors already built into the arm. They collected a gesture dataset, then compared how different data representations and model shapes affected accuracy. The key finding: turning joint sensor readings into spectrograms — visual frequency maps similar to those used in audio processing — drove accuracy above 95% for both contact detection and gesture classification. The specific CNN architecture mattered far less than that representational choice.
Most human-robot collaboration research leans on cameras or sensor-laden robot "skins" to detect touch, both of which add cost and complexity. Getting comparable results from sensors a robot already has changes the economics of deployment — a factory arm doesn't need a retrofit before a human worker can tap it to pause a task.
The 95% figure holds up when the robot moves to new poses it wasn't trained on, which is the test that usually breaks gesture systems in practice. Whether it survives the noise of a real factory floor — vibration, tool contact, incidental bumps — is the next question the lab setting can't answer.