Imitation learning often relies on low-cost passive leader arms for scalable data collection, but these arms have no restoring forces, so redundant joints drift and produce highly inconsistent trajectories. The authors propose a simple hardware fix: a passive joint module combining an adjustable mechanical clutch with a 3D-printed spiral spring that pulls each joint toward a user-defined neutral posture.
This passive restoring force acts as a physical prior, equivalent to joint-space L2 regularization—what the authors call "Hardware-as-Regularization." In experiments, the modified arm reduced trajectory variance substantially compared with a naive passive arm, preserved movement smoothness, and completed tasks faster. Its trajectory consistency matched that of an active virtual spring from the same actuator class.
Downstream success rates improved markedly: on Block Stacking, the regularized arm achieved 73.3% success versus 36.7% for the unregularized arm. A stiffness sweep showed that consistency degrades monotonically as stiffness decreases, confirming that mechanical stiffness behaves like a regularization weight. The work suggests that carefully chosen passive mechanics can substitute for algorithmic regularization in learning from demonstration.