Kinematic Neural Networks: A New Paradigm for Embodied Mechanical Intelligence
Abstract
Physical Artificial Intelligence and Mechanical Neural Networks (MNNs) embed information processing in physical structures. This paper investigates a complementary route based on deterministic rigid-body kinematics. We formulate a Kinematic Neural Network (KNN) as a layered linkage computation graph whose geometric parameters are optimized end-to-end by gradient descent. Explicit planar constraint equations, assembly-branch selection, and feasibility conditions are given for the fundamental units. Compact digit and letter tasks are used as multi-input mechanical benchmarks, and an additively manufactured prototype verifies physical feasibility while revealing a simulation-to-hardware gap. An application-oriented design study maps continuous $4\times 4$ foot-contact displacement fields to support-locking (SL), compliance-release (CR), and buffer/warning (BW) output states as labels for prospective downstream functions. Because contact input, geometric propagation, and output-state generation are all mechanical displacements, the proposed path evolves continuously with contact and does not require electrical sampling, communication, or digital inference. The optimized 30-unit, 60-link design achieves 99.53% accuracy on 1500 held-out simulated fields without reachability or slot-limit violations. A fully mechanical $4\times 4$ prototype achieves 45/45 correct output-region assignments across five repeated trials on each of nine terrain configurations; this hardware study is limited to contact-to-output displacement correspondence.
