Transformer-Based Identification of Open-Chain Robots

Transformer-Based Identification of Open-Chain Robot Kinematics from Multisignal Windows

This paper studies the identification of unknown open-chain

robot morphologies from time windows containing commanded

references, measured joint states, and end-effector observations. The

estimator predicts the number of joints, the revolute/prismatic sequence,

the Product of Exponentials (POE) screw coordinates, and the home

pose of the end effector. The proposed model is a temporal transformer

encoder with multitask heads and a hybrid objective that combines

discrete classification, masked continuous regression, pose-consistency,

and screw regularization. POE plus home pose is adopted instead

of direct Denavit-Hartenberg regression because it provides a more

stable and physically meaningful target for variable-DOF serial chains.

Simulation results on a noisy benchmark of serial manipulators support

the feasibility of the proposed approach for robot self-modeling and

kinematic identification.