Abstract
<title>Abstract</title> <p> The scaling of embodied intelligence is fundamentally constrained by a <bold>physical bottleneck:</bold> human supervision must be channeled through rigid, morphology-locked master-slave manipulators, rendering policies non-transferable across robotic platforms. Here we propose and experimentally validate a <bold>Kinematic Decoupling Control Theory</bold> , demonstrating that human manipulation intent—specifically, rotational orientation versus spatial translation—can be strictly orthogonally decomposed in information space, entirely eliminating the requirement for physical kinesthetic twins. We instantiate this theory through a dematerialized teleoperation framework (DM-Teleop) utilizing immersive VR and adaptive Jacobian decoupling. Our work establishes three paradigm-shifting results. <bold>First (Theoretical Novelty)</bold> , we formalize and experimentally validate the subspace orthogonality condition that the traditional coupled damped least-squares (DLS) Jacobian suffers from inherent ill-conditioning that inextricably couples wrist rotation with elbow translation; our sequential sub-space projection solver suppresses end-effector drift during pure rotation from the physical limit of 18.2±4.1 mm to 1.2±0.3 mm—surpassing the human haptic threshold by sixfold. <bold>Second (Performance Excellence)</bold> , we elevate teleoperation from probabilistic success to deterministic control: untrained operators achieve 100% task success in pick-and-place operations, with a NASA-TLX workload score (31.5) statistically indistinguishable from physical master arms (p0.05), while adaptive damping suppresses singularity-induced joint jerks by a factor of <bold>4.6. Third (Broad Impact)</bold> , we demonstrate <bold>zero-shot morphological generalization</bold> across 6-DOF and 7-DOF architectures without any model retraining, and achieve <bold>100% jitter-free online intervention</bold> during autonomous policy rollout with 85±12 ms takeover latency. Our findings overturn a four-decade-old robotics dogma—that kinesthetic force feedback is indispensable for intuitive control. We show that <bold>multi-modal sensory substitution</bold> (quasi-visual overlays + encoded vibrotactile risk maps) reconstructs perceived object stiffness with 92% accuracy, functionally surpassing physical force fidelity. This framework constitutes a universal API between foundation models and physical embodiments, paving the way for "teach once, deploy everywhere" in general-purpose robotics. </p>