A Pairing-Free End-to-End Partial Supervised Method for Mapping Dexterous Postures Between Human and Robot Hands

Synergy-based end-to-end mapping methods for dissimilar hand structures combine the intuitive control of end-to-end learning with the generalization capabilities of synergy-based representations. However, low-dimensional synergy representations cannot capture all hand postures, particularly those requiring precise and independent finger movements. Although effective at mapping simple hand postures without paired cross-hand training data, the original End-to-End Partial Supervised Mapping (EPSM) method is limited in its ability to represent complex precision postures. This work expands EPSM by learning additional latent features that complement the original synergy-informed features, allowing the representation of more complex postures across heterogeneous hand structures. The Expanded EPSM framework is evaluated using four robotic hand configurations with three, four, and five fingers and differing joint arrangements. The results demonstrate that the expanded framework preserves the established synergy features while learning complementary features that improve precision-posture mapping across heterogeneous robotic hands without requiring paired cross-hand training data.